diff --git a/data/links.yml b/data/links.yml
index 2f54cb4..013746e 100644
--- a/data/links.yml
+++ b/data/links.yml
@@ -25,7 +25,7 @@ links:
   gemma-terms:
     url: /resources/terms
   prohibited-use:
-    url: /resources/prohibited_use_policy
+    url: /resources/prohibited-use
 
   # --- Weights and collections -----------------------------------------------------------
   hf-gemma4:
@@ -122,7 +122,7 @@ links:
   gemmaverse:
     url: https://deepmind.google/models/gemma/gemmaverse/
   gemma-cookbook:
-    url: https://github.com/google-gemma/gemma-cookbook
+    url: https://github.com/google-gemma/cookbook
   gemma-skills-repo:
     url: https://github.com/google-gemma/gemma-skills
   awesome-gemma:
diff --git a/data/models.yml b/data/models.yml
index c5c12dd..c22dab7 100644
--- a/data/models.yml
+++ b/data/models.yml
@@ -304,8 +304,10 @@ gemma4:
         hf: google/gemma-4-E2B-it
         hf_base: google/gemma-4-E2B
         hf_qat: google/gemma-4-E2B-it-qat-w4a16-ct
-        gguf: google/gemma-4-E2B-it-qat-q4_0-gguf
+        gguf: ggml-org/gemma-4-E2B-it-GGUF
         gguf_qat: google/gemma-4-E2B-it-qat-q4_0-gguf
+        gguf_file: gemma-4-E2B-it-Q4_0.gguf
+        gguf_mmproj: mmproj-gemma-4-E2B-it-Q8_0.gguf
         ollama: gemma4:e2b
         mlx: mlx-community/gemma-4-E2B-it-qat-4bit
         lmstudio: google/gemma-4-e2b
@@ -313,6 +315,7 @@ gemma4:
         keras_base: gemma4_2b
         keras_assistant: gemma4_instruct_2b_assistant
         litert: litert-community/gemma-4-E2B-it-litert-lm
+        litert_file: gemma-4-E2B-it.litertlm
       memory:                          # GB, static weights + ~20% overhead
         bf16: 11.4
         sfp8: 5.7
@@ -335,8 +338,10 @@ gemma4:
         hf: google/gemma-4-E4B-it
         hf_base: google/gemma-4-E4B
         hf_qat: google/gemma-4-E4B-it-qat-w4a16-ct
-        gguf: google/gemma-4-E4B-it-qat-q4_0-gguf
+        gguf: ggml-org/gemma-4-E4B-it-GGUF
         gguf_qat: google/gemma-4-E4B-it-qat-q4_0-gguf
+        gguf_file: gemma-4-E4B-it-Q4_0.gguf
+        gguf_mmproj: mmproj-gemma-4-E4B-it-Q8_0.gguf
         ollama: gemma4:e4b
         mlx: mlx-community/gemma-4-E4B-it-qat-4bit
         lmstudio: google/gemma-4-e4b
@@ -344,6 +349,7 @@ gemma4:
         keras_base: gemma4_4b
         keras_assistant: gemma4_instruct_4b_assistant
         litert: litert-community/gemma-4-E4B-it-litert-lm
+        litert_file: gemma-4-E4B-it.litertlm
       memory:
         bf16: 17.9
         sfp8: 8.9
@@ -366,8 +372,10 @@ gemma4:
         hf: google/gemma-4-12B-it
         hf_base: google/gemma-4-12B
         hf_qat: google/gemma-4-12B-it-qat-w4a16-ct
-        gguf: google/gemma-4-12B-it-qat-q4_0-gguf
+        gguf: ggml-org/gemma-4-12B-it-GGUF
         gguf_qat: google/gemma-4-12B-it-qat-q4_0-gguf
+        gguf_file: gemma-4-12B-it-Q4_0.gguf
+        gguf_mmproj: mmproj-gemma-4-12B-it-Q8_0.gguf
         ollama: gemma4:12b
         mlx: mlx-community/gemma-4-12B-it-qat-4bit
         lmstudio: google/gemma-4-12b
@@ -391,8 +399,10 @@ gemma4:
       ids:                             # verified repo/tag strings; see PLAN.md
         hf: google/gemma-4-26B-A4B-it
         hf_base: google/gemma-4-26B-A4B
-        gguf: google/gemma-4-26B-A4B-it-qat-q4_0-gguf
+        gguf: ggml-org/gemma-4-26B-A4B-it-GGUF
         gguf_qat: google/gemma-4-26B-A4B-it-qat-q4_0-gguf
+        gguf_file: gemma-4-26B-A4B-it-Q4_0.gguf
+        gguf_mmproj: mmproj-gemma-4-26B-A4B-it-Q8_0.gguf
         ollama: gemma4:26b
         mlx: mlx-community/gemma-4-26B-A4B-it-qat-4bit
         lmstudio: google/gemma-4-26b-a4b
@@ -420,8 +430,10 @@ gemma4:
         hf: google/gemma-4-31B-it
         hf_base: google/gemma-4-31B
         hf_qat: google/gemma-4-31B-it-qat-w4a16-ct
-        gguf: google/gemma-4-31B-it-qat-q4_0-gguf
+        gguf: ggml-org/gemma-4-31B-it-GGUF
         gguf_qat: google/gemma-4-31B-it-qat-q4_0-gguf
+        gguf_file: gemma-4-31B-it-Q4_0.gguf
+        gguf_mmproj: mmproj-gemma-4-31B-it-Q8_0.gguf
         ollama: gemma4:31b
         mlx: mlx-community/gemma-4-31B-it-qat-4bit
         lmstudio: google/gemma-4-31b
diff --git a/docs/deployment/huggingface.md b/docs/deployment/huggingface.md
index 45c6cc7..7d9fd4f 100644
--- a/docs/deployment/huggingface.md
+++ b/docs/deployment/huggingface.md
@@ -12,7 +12,7 @@ colab: https://colab.research.google.com/github/google-gemma/docs/blob/main/docs
 Generating text, summarizing, and analysing content are just some of the tasks you can accomplish with Gemma open models. This tutorial shows you how to get started running Gemma using Hugging Face Transformers using both text and image input to generate text content. The Transformers Python library provides a API for accessing pre-trained generative AI models, including Gemma. For more information, see the [Transformers](https://huggingface.co/docs/transformers/en/index) documentation.
 
 !!! Note
-    Gemma 3 and later models support input of both images and text. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../get-started/index.md).
+    Gemma 3 and later models support input of both images and text. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../models/paligemma/index.md).
 
 
 ### Install Python packages
@@ -143,6 +143,7 @@ You can include multiple images in your prompt by including additional `"type":
     Do not use `<|image|>`{.token-multi}, `<start_of_image>`{.token-multi} or `<image_soft_token>`{.token-multi} tokens in text portion of a prompt template as this approach creates redundant tokens and processing errors.
     
 
+<!-- notebook:requires-start audio -->
 ## Generate text from audio data
 
 With [Gemma 4](../models/gemma-4/index.md) and [Gemma 3n](../models/gemma-3n/index.md), you can use audio data as part of your prompt. This section shows how to use the Transformers library to load and configure a Gemma model to use audio data and text input to generate text output.
@@ -151,7 +152,7 @@ With [Gemma 4](../models/gemma-4/index.md) and [Gemma 3n](../models/gemma-3n/ind
 
 When generating content with audio, use a prompt template to structure your request. A prompt template allows you to specify input from specific roles, such as `user`{.token-role} or `model`{.token-role}, and is a required format for managing multi-turn chat interactions with Gemma models. The following example code shows how to constuct a prompt template for Gemma with audio data input:
 
-```python { .notebook-cell }
+```python { .notebook-cell data-requires="audio" }
 from transformers import GenerationConfig
 config = GenerationConfig.from_pretrained(MODEL_ID)
 config.max_new_tokens = 512
@@ -183,6 +184,7 @@ You can include multiple audio files in your prompt by including additional `"ty
 !!! Note
     Do not use `<|audio|>`{.token-multi} or `<audio_soft_token>`{.token-multi} tokens in text portion of a prompt template as this approach creates redundant tokens and processing errors.
     
+<!-- notebook:requires-end -->
 
 ## Next steps
 
diff --git a/docs/deployment/keras.md b/docs/deployment/keras.md
index a5ff282..8278b98 100644
--- a/docs/deployment/keras.md
+++ b/docs/deployment/keras.md
@@ -11,7 +11,7 @@ colab: https://colab.research.google.com/github/google-gemma/docs/blob/main/docs
 
 Generating content, summarizing, and analysing content are just some of the tasks you can accomplish with Gemma open models. This tutorial shows you how to get started running Gemma using Keras, including generating text content with text and image input. [Keras](https://keras.io/) provides implementations for running Gemma and other models using JAX, PyTorch, and TensorFlow. If you're new to Keras, you might want to read [Getting started with Keras](https://keras.io/getting_started/) before you begin.
 
-Gemma 3 and later models support text and image input. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../get-started/index.md).
+Gemma 3 and later models support text and image input. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../models/paligemma/index.md).
 !!! Note
     This notebook won't run on T4 if you try with Gemma 4. Use L4 or higher GPUs.
     
diff --git a/docs/deployment/llamacpp.md b/docs/deployment/llamacpp.md
index 781101f..1c93499 100644
--- a/docs/deployment/llamacpp.md
+++ b/docs/deployment/llamacpp.md
@@ -13,41 +13,29 @@ most capable local runtime for Gemma 4 — **image, audio and video input all wo
 
 ## Install
 
-=== "macOS / Linux"
+=== "curl"
 
     ```bash
-    brew install llama.cpp
+    curl -LsSf https://llama.app/install.sh | sh
     ```
 
-=== "Windows"
+=== "powershell"
 
-    ```bash
-    winget install llama.cpp
-    ```
-
-=== "conda"
-
-    ```bash
-    conda install -c conda-forge llama.cpp
-    ```
-
-=== "Nix"
-
-    ```bash
-    nix profile install nixpkgs#llama-cpp
+    ```powershell
+    irm https://llama.app/install.ps1 | iex
     ```
 
 ## Run
 
 ```bash { .model-swap }
-llama-cli -hf ggml-org/gemma-4-E2B-it-GGUF --prompt "Write a poem about the Kraken."
+llama cli -hf ggml-org/gemma-4-E2B-it-GGUF --prompt "Write a poem about the Kraken."
 ```
 
 `-hf` downloads the weights on first use, including the multimodal projector. With a system
 prompt:
 
 ```bash { .model-swap }
-llama-cli -hf ggml-org/gemma-4-E2B-it-GGUF -sys "You are Hong Gildong." -p "Who are you?"
+llama cli -hf ggml-org/gemma-4-E2B-it-GGUF -sys "You are Hong Gildong." -p "Who are you?"
 ```
 
 ## Model repositories
@@ -64,7 +52,7 @@ For 4-bit, prefer Google's [QAT](../get-started/qat.md) weights — trained with
 than compressed afterwards:
 
 ```bash { .model-swap }
-llama-cli -hf google/gemma-4-12B-it-qat-q4_0-gguf -p "Explain PLE in one paragraph"
+llama cli -hf google/gemma-4-E2B-it-qat-q4_0-gguf -p "Explain gravity in one paragraph"
 ```
 
 Available as `google/gemma-4-{E2B,E4B,12B,26B-A4B,31B}-it-qat-q4_0-gguf`.
@@ -72,13 +60,13 @@ Available as `google/gemma-4-{E2B,E4B,12B,26B-A4B,31B}-it-qat-q4_0-gguf`.
 ## Multimodal
 
 ```bash { .model-swap }
-llama-mtmd-cli -hf ggml-org/gemma-4-E2B-it-GGUF
+llama cli -hf ggml-org/gemma-4-E2B-it-GGUF --image "image.png" -p "Describe this image."
 ```
 
 With local files, pass the projector explicitly:
 
-```bash
-llama-server -m gemma-4-12B-it-Q4_K_M.gguf --mmproj mmproj-gemma-4-12B-it.gguf
+```bash { .model-swap }
+llama cli -m gemma-4-E2B-it-Q4_0.gguf --mm mmproj-gemma-4-E2B-it-Q8_0.gguf
 ```
 
 !!! note
@@ -88,7 +76,7 @@ llama-server -m gemma-4-12B-it-Q4_K_M.gguf --mmproj mmproj-gemma-4-12B-it.gguf
 ## Server
 
 ```bash { .model-swap }
-llama-server -hf ggml-org/gemma-4-E2B-it-GGUF
+llama server -hf ggml-org/gemma-4-E2B-it-GGUF
 ```
 
 Browser UI at `http://localhost:8080`, OpenAI-compatible API at `http://localhost:8080/v1`.
diff --git a/docs/deployment/unsloth.md b/docs/deployment/unsloth.md
index f527525..65280f7 100644
--- a/docs/deployment/unsloth.md
+++ b/docs/deployment/unsloth.md
@@ -86,17 +86,19 @@ and both GGUF and MLX files are supported.
 
 ## llama.cpp
 
-Build `llama.cpp` first — set `-DGGML_CUDA=OFF` for CPU-only machines and for Apple silicon,
-where Metal is on by default:
+A few options to get `llama.cpp` installed on your machine:
 
-```bash
-git clone https://github.com/ggml-org/llama.cpp
-cmake llama.cpp -B llama.cpp/build \
-    -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=ON
-cmake --build llama.cpp/build --config Release -j --clean-first \
-    --target llama-cli llama-mtmd-cli llama-server llama-gguf-split
-cp llama.cpp/build/bin/llama-* llama.cpp
-```
+=== "curl"
+
+    ```bash
+    curl -LsSf https://llama.app/install.sh | sh
+    ```
+
+=== "powershell"
+
+    ```powershell
+    irm https://llama.app/install.ps1 | iex
+    ```
 
 Then pull and run a model straight from Hugging Face. There is no need to set a context length —
 llama.cpp uses exactly what is required.
@@ -104,8 +106,7 @@ llama.cpp uses exactly what is required.
 === "E2B"
 
     ```bash
-    export LLAMA_CACHE="unsloth/gemma-4-E2B-it-GGUF"
-    ./llama.cpp/llama-cli \
+    llama cli \
         -hf unsloth/gemma-4-E2B-it-GGUF:Q8_0 \
         --temp 1.0 --top-p 0.95 --top-k 64
     ```
@@ -113,8 +114,7 @@ llama.cpp uses exactly what is required.
 === "E4B"
 
     ```bash
-    export LLAMA_CACHE="unsloth/gemma-4-E4B-it-GGUF"
-    ./llama.cpp/llama-cli \
+    llama cli \
         -hf unsloth/gemma-4-E4B-it-GGUF:Q8_0 \
         --temp 1.0 --top-p 0.95 --top-k 64
     ```
@@ -122,8 +122,7 @@ llama.cpp uses exactly what is required.
 === "12B"
 
     ```bash
-    export LLAMA_CACHE="unsloth/gemma-4-12B-it-GGUF"
-    ./llama.cpp/llama-cli \
+    llama cli \
         -hf unsloth/gemma-4-12b-it-GGUF:UD-Q4_K_XL \
         --temp 1.0 --top-p 0.95 --top-k 64
     ```
@@ -131,8 +130,7 @@ llama.cpp uses exactly what is required.
 === "26B A4B"
 
     ```bash
-    export LLAMA_CACHE="unsloth/gemma-4-26B-A4B-it-GGUF"
-    ./llama.cpp/llama-cli \
+    llama cli \
         -hf unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_XL \
         --temp 1.0 --top-p 0.95 --top-k 64
     ```
@@ -140,8 +138,7 @@ llama.cpp uses exactly what is required.
 === "31B"
 
     ```bash
-    export LLAMA_CACHE="unsloth/gemma-4-31B-it-GGUF"
-    ./llama.cpp/llama-cli \
+    llama cli \
         -hf unsloth/gemma-4-31B-it-GGUF:UD-Q4_K_XL \
         --temp 1.0 --top-p 0.95 --top-k 64
     ```
@@ -158,7 +155,7 @@ hf download unsloth/gemma-4-26B-A4B-it-GGUF \
 ```
 
 ```bash
-./llama.cpp/llama-cli \
+llama cli \
     --model unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf \
     --mmproj unsloth/gemma-4-26B-A4B-it-GGUF/mmproj-BF16.gguf \
     --temp 1.0 --top-p 0.95 --top-k 64
@@ -167,7 +164,7 @@ hf download unsloth/gemma-4-26B-A4B-it-GGUF \
 ### Server
 
 ```bash
-./llama.cpp/llama-server \
+llama server \
     --model unsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf \
     --mmproj unsloth/gemma-4-26B-A4B-it-GGUF/mmproj-BF16.gguf \
     --temp 1.0 --top-p 0.95 --top-k 64 \
diff --git a/docs/get-started/capabilities/function-calling.md b/docs/get-started/capabilities/function-calling.md
index 6aa8366..4bcedf9 100644
--- a/docs/get-started/capabilities/function-calling.md
+++ b/docs/get-started/capabilities/function-calling.md
@@ -264,7 +264,7 @@ if calls:
 !!! Note
     For optimal results, append the tool execution result to your message history using the specific format below. This ensures the chat template correctly generates the required token structure (e.g., `response:get_current_weather{temperature:15,weather:<|"|>sunny<|"|>}`).
 
-```python { .notebook-cell }
+```python
 "tool_responses": [
   {
     "name": function_name,
@@ -275,7 +275,7 @@ if calls:
 
 In case of multiple independent requests:
 
-```python { .notebook-cell }
+```python
 "tool_responses": [
   {
     "name": function_name_1,
diff --git a/docs/get-started/capabilities/image.md b/docs/get-started/capabilities/image.md
index 5e72819..8ffc8ab 100644
--- a/docs/get-started/capabilities/image.md
+++ b/docs/get-started/capabilities/image.md
@@ -320,16 +320,16 @@ output = vqa_pipe(messages, return_full_text=False, generate_kwargs=gen_kwargs)
 print(output[0]['generated_text'])
 
 draw_results(output[0]['generated_text'])
-``` 
+```
 
-```{ .no-copy .output }
+```` { .no-copy .output }
 ```json
 [
   {"box_2d": [243, 252, 956, 415], "label": "person"},
   {"box_2d": [356, 606, 654, 802], "label": "cat"}
 ]
 ```<turn|>
-```
+````
 
 ![png](../../assets/capabilities/output_b041486b_1.png)
 
@@ -409,14 +409,14 @@ print(output[0]['generated_text'])
 draw_results(output[0]['generated_text'])
 ```
 
-```{ .no-copy .output }
+```` { .no-copy .output }
 ```json
 [
   {"box_2d": [454, 871, 533, 923], "label": "person"},
   {"box_2d": [551, 316, 665, 628], "label": "car"}
 ]
 ```<turn|>
-```
+````
 
 ![png](../../assets/capabilities/output_bc2920e6_1.png)
 
@@ -506,7 +506,7 @@ plt.tight_layout()
 plt.show()
 ```
 
-```{ .no-copy .output }
+```` { .no-copy .output }
 Budget Size: 70
 # of Image Tokens: 64
 [transformers] You seem to be using the pipelines sequentially on GPU. In order to maximize efficiency please use a dataset
@@ -551,7 +551,7 @@ Budget Size: 560
   {"box_2d": [404, 867, 520, 925], "label": "person"}
 ]
 ```<turn|>
-```
+````
 
 ![png](../../assets/capabilities/output_76f17518_3.png)
 
diff --git a/docs/get-started/index.md b/docs/get-started/index.md
index 5937f90..29d8865 100644
--- a/docs/get-started/index.md
+++ b/docs/get-started/index.md
@@ -40,7 +40,7 @@ and move up without rewriting anything.
 
     pipe = pipeline(
         "text-generation",
-        model="google/gemma-4-12B-it",
+        model="google/gemma-4-E2B-it",
         device_map="auto"
     )
     output = pipe("Explain quantum computing simply:", max_new_tokens=150)
@@ -51,14 +51,14 @@ and move up without rewriting anything.
     Install [`llama.cpp`](https://llama.app/), then run:
 
     ```bash { .model-swap }
-    llama-cli -hf google/gemma-4-12B-it-qat-q4_0-gguf --prompt "Write a poem about the Kraken."
+    llama cli -hf google/gemma-4-E2B-it-qat-q4_0-gguf --prompt "Write a poem about the Kraken."
     ```
 
 === "Ollama"
     Install [`ollama`](https://ollama.com/download), then run:
 
     ```bash { .model-swap }
-    ollama run gemma4:12b "Explain quantum computing simply:"
+    ollama run gemma4:e2b "Explain quantum computing simply:"
     ```
 
 === "Apple MLX"
@@ -66,7 +66,7 @@ and move up without rewriting anything.
     Install [`MLX LM`](https://github.com/ml-explore/mlx-lm), then run:
 
     ```bash { .model-swap }
-    mlx_lm.generate --model mlx-community/gemma-4-12B-it-qat-4bit \
+    mlx_lm.generate --model mlx-community/gemma-4-E2B-it-qat-4bit \
         --prompt "Explain quantum computing simply:"
     ```
 
@@ -75,14 +75,14 @@ and move up without rewriting anything.
     Install [`vLLM`](https://docs.vllm.ai/en/latest/getting_started/installation/), then run:
 
     ```bash { .model-swap }
-    vllm serve google/gemma-4-26B-A4B-it
+    vllm serve google/gemma-4-E2B-it
     ```
 
 === "LMStudio"
     Install [`LMStudio`](https://lmstudio.ai/), then run:
 
     ```bash { .model-swap }
-    lms chat google/gemma-4-12b
+    lms chat google/gemma-4-e2b
     ```
 
 !!! example "Using a coding agent?"
diff --git a/docs/get-started/qat.md b/docs/get-started/qat.md
index 66de8a3..e4489e2 100644
--- a/docs/get-started/qat.md
+++ b/docs/get-started/qat.md
@@ -29,7 +29,7 @@ Names follow `google/gemma-4-{SIZE}-it{SUFFIX}`, for example:
 
 ```bash { .model-swap }
 # Local, via llama.cpp
-llama-cli -hf google/gemma-4-12B-it-qat-q4_0-gguf
+llama cli -hf google/gemma-4-12B-it-qat-q4_0-gguf
 ```
 
 ```bash { .model-fixed }
diff --git a/docs/javascripts/model-data.js b/docs/javascripts/model-data.js
index 25f3798..8d69455 100644
--- a/docs/javascripts/model-data.js
+++ b/docs/javascripts/model-data.js
@@ -31,15 +31,18 @@ window.GEMMA_MODELS = {
       "hf": "google/gemma-4-E2B-it",
       "hf_base": "google/gemma-4-E2B",
       "hf_qat": "google/gemma-4-E2B-it-qat-w4a16-ct",
-      "gguf": "google/gemma-4-E2B-it-qat-q4_0-gguf",
+      "gguf": "ggml-org/gemma-4-E2B-it-GGUF",
       "gguf_qat": "google/gemma-4-E2B-it-qat-q4_0-gguf",
+      "gguf_file": "gemma-4-E2B-it-Q4_0.gguf",
+      "gguf_mmproj": "mmproj-gemma-4-E2B-it-Q8_0.gguf",
       "ollama": "gemma4:e2b",
       "mlx": "mlx-community/gemma-4-E2B-it-qat-4bit",
       "lmstudio": "google/gemma-4-e2b",
       "keras": "gemma4_instruct_2b",
       "keras_base": "gemma4_2b",
       "keras_assistant": "gemma4_instruct_2b_assistant",
-      "litert": "litert-community/gemma-4-E2B-it-litert-lm"
+      "litert": "litert-community/gemma-4-E2B-it-litert-lm",
+      "litert_file": "gemma-4-E2B-it.litertlm"
     },
     "inputs": [
       "text",
@@ -57,15 +60,18 @@ window.GEMMA_MODELS = {
       "hf": "google/gemma-4-E4B-it",
       "hf_base": "google/gemma-4-E4B",
       "hf_qat": "google/gemma-4-E4B-it-qat-w4a16-ct",
-      "gguf": "google/gemma-4-E4B-it-qat-q4_0-gguf",
+      "gguf": "ggml-org/gemma-4-E4B-it-GGUF",
       "gguf_qat": "google/gemma-4-E4B-it-qat-q4_0-gguf",
+      "gguf_file": "gemma-4-E4B-it-Q4_0.gguf",
+      "gguf_mmproj": "mmproj-gemma-4-E4B-it-Q8_0.gguf",
       "ollama": "gemma4:e4b",
       "mlx": "mlx-community/gemma-4-E4B-it-qat-4bit",
       "lmstudio": "google/gemma-4-e4b",
       "keras": "gemma4_instruct_4b",
       "keras_base": "gemma4_4b",
       "keras_assistant": "gemma4_instruct_4b_assistant",
-      "litert": "litert-community/gemma-4-E4B-it-litert-lm"
+      "litert": "litert-community/gemma-4-E4B-it-litert-lm",
+      "litert_file": "gemma-4-E4B-it.litertlm"
     },
     "inputs": [
       "text",
@@ -83,8 +89,10 @@ window.GEMMA_MODELS = {
       "hf": "google/gemma-4-12B-it",
       "hf_base": "google/gemma-4-12B",
       "hf_qat": "google/gemma-4-12B-it-qat-w4a16-ct",
-      "gguf": "google/gemma-4-12B-it-qat-q4_0-gguf",
+      "gguf": "ggml-org/gemma-4-12B-it-GGUF",
       "gguf_qat": "google/gemma-4-12B-it-qat-q4_0-gguf",
+      "gguf_file": "gemma-4-12B-it-Q4_0.gguf",
+      "gguf_mmproj": "mmproj-gemma-4-12B-it-Q8_0.gguf",
       "ollama": "gemma4:12b",
       "mlx": "mlx-community/gemma-4-12B-it-qat-4bit",
       "lmstudio": "google/gemma-4-12b"
@@ -104,8 +112,10 @@ window.GEMMA_MODELS = {
     "ids": {
       "hf": "google/gemma-4-26B-A4B-it",
       "hf_base": "google/gemma-4-26B-A4B",
-      "gguf": "google/gemma-4-26B-A4B-it-qat-q4_0-gguf",
+      "gguf": "ggml-org/gemma-4-26B-A4B-it-GGUF",
       "gguf_qat": "google/gemma-4-26B-A4B-it-qat-q4_0-gguf",
+      "gguf_file": "gemma-4-26B-A4B-it-Q4_0.gguf",
+      "gguf_mmproj": "mmproj-gemma-4-26B-A4B-it-Q8_0.gguf",
       "ollama": "gemma4:26b",
       "mlx": "mlx-community/gemma-4-26B-A4B-it-qat-4bit",
       "lmstudio": "google/gemma-4-26b-a4b",
@@ -129,8 +139,10 @@ window.GEMMA_MODELS = {
       "hf": "google/gemma-4-31B-it",
       "hf_base": "google/gemma-4-31B",
       "hf_qat": "google/gemma-4-31B-it-qat-w4a16-ct",
-      "gguf": "google/gemma-4-31B-it-qat-q4_0-gguf",
+      "gguf": "ggml-org/gemma-4-31B-it-GGUF",
       "gguf_qat": "google/gemma-4-31B-it-qat-q4_0-gguf",
+      "gguf_file": "gemma-4-31B-it-Q4_0.gguf",
+      "gguf_mmproj": "mmproj-gemma-4-31B-it-Q8_0.gguf",
       "ollama": "gemma4:31b",
       "mlx": "mlx-community/gemma-4-31B-it-qat-4bit",
       "lmstudio": "google/gemma-4-31b",
diff --git a/docs/javascripts/model-switcher.js b/docs/javascripts/model-switcher.js
index dd4209d..f4def15 100644
--- a/docs/javascripts/model-switcher.js
+++ b/docs/javascripts/model-switcher.js
@@ -29,6 +29,11 @@
     }
   }
   index.sort((a, b) => b.text.length - a.text.length);
+  const byText = new Map(index.map((entry) => [entry.text, entry]));
+  const escapeRegExp = (str) => str.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+  const pattern = index.length
+    ? new RegExp(index.map((entry) => escapeRegExp(entry.text)).join("|"), "g")
+    : null;
 
   const blocks = () => document.querySelectorAll(".model-swap code, code.model-swap");
   const displayName = (id) => models[id]?.fullName || models[id]?.name || id;
@@ -59,9 +64,11 @@
 
   const enginesInUse = () => {
     const found = new Set();
+    if (!pattern) return found;
     for (const block of blocks()) {
-      for (const entry of index) {
-        if (block.textContent.includes(entry.text)) found.add(entry.engine);
+      for (const [match] of block.textContent.matchAll(pattern)) {
+        const entry = byText.get(match);
+        if (entry) found.add(entry.engine);
       }
     }
     return found;
@@ -78,26 +85,74 @@
   };
 
   const detect = () => {
+    if (!pattern) return null;
     for (const block of blocks()) {
-      const hit = index.find((entry) => block.textContent.includes(entry.text));
-      if (hit) return hit.model;
+      for (const [match] of block.textContent.matchAll(pattern)) {
+        const entry = byText.get(match);
+        if (entry) return entry.model;
+      }
     }
     return null;
   };
 
   const rewriteNode = (node, target) => {
+    if (!pattern) return;
     const original = node.__gemmaOriginal ?? node.nodeValue;
-    let next = original;
-    for (const entry of index) {
-      if (!next.includes(entry.text)) continue;
-      const replacement = models[target]?.ids?.[entry.engine];
-      if (replacement) next = next.split(entry.text).join(replacement);
-    }
+    const next = original.replace(pattern, (match) => {
+      const entry = byText.get(match);
+      return (entry && models[target]?.ids?.[entry.engine]) || match;
+    });
     if (next === node.nodeValue) return;
     node.__gemmaOriginal = original;
     node.nodeValue = next;
   };
 
+  const applyRequires = (target) => {
+    const inputs = models[target]?.inputs || [];
+    const article = document.querySelector(".md-content__inner");
+    if (article) {
+      const walker = document.createTreeWalker(article, NodeFilter.SHOW_COMMENT);
+      const startRe = /^\s*notebook:requires-start\s+([a-z0-9_-]+)\s*$/i;
+      const endRe = /^\s*notebook:requires-end\s*$/i;
+      const starts = [];
+      let comment;
+      while ((comment = walker.nextNode())) {
+        const match = comment.nodeValue.match(startRe);
+        if (match) starts.push({ node: comment, modality: match[1].toLowerCase() });
+      }
+      for (const { node, modality } of starts) {
+        const supported = inputs.includes(modality);
+        let cur = node.nextSibling;
+        while (cur) {
+          if (cur.nodeType === Node.COMMENT_NODE && endRe.test(cur.nodeValue)) break;
+          if (cur.nodeType === Node.ELEMENT_NODE) {
+            cur.hidden = !supported;
+            cur.style.display = supported ? "" : "none";
+            if (cur.id && /^H[1-6]$/.test(cur.tagName)) {
+              const targetHash = `#${cur.id}`;
+              for (const link of document.querySelectorAll(".md-nav--secondary a.md-nav__link")) {
+                if (link.hash !== targetHash) continue;
+                const item = link.closest(".md-nav__item");
+                if (item) {
+                  item.hidden = !supported;
+                  item.style.display = supported ? "" : "none";
+                }
+              }
+            }
+          }
+          cur = cur.nextSibling;
+        }
+      }
+    }
+    for (const el of document.querySelectorAll("[data-requires]:not(.page-context)")) {
+      const modality = el.dataset.requires?.trim().toLowerCase();
+      if (!modality) continue;
+      const supported = inputs.includes(modality);
+      el.hidden = !supported;
+      el.style.display = supported ? "" : "none";
+    }
+  };
+
   const apply = (target) => {
     if (!models[target]) return;
     for (const block of blocks()) {
@@ -105,6 +160,31 @@
       let node;
       while ((node = walker.nextNode())) rewriteNode(node, target);
     }
+    applyRequires(target);
+  };
+
+  const positionMenu = (picker) => {
+    const menu = picker.querySelector(".page-context__picker-menu");
+    const strip = picker.closest(".page-context");
+    if (!menu || menu.hidden || !strip) return;
+    menu.style.left = "";
+    menu.style.right = "";
+    const stripRect = strip.getBoundingClientRect();
+    const pickerRect = picker.getBoundingClientRect();
+    const menuRect = menu.getBoundingClientRect();
+    if (menuRect.right > stripRect.right) {
+      if (pickerRect.right - menuRect.width >= stripRect.left) {
+        menu.style.left = "auto";
+        menu.style.right = "0";
+      } else {
+        const offset = Math.max(
+          stripRect.left - pickerRect.left,
+          stripRect.right - pickerRect.left - menuRect.width
+        );
+        menu.style.left = `${offset}px`;
+        menu.style.right = "auto";
+      }
+    }
   };
 
   const closePicker = (picker, returnFocus = false) => {
@@ -209,6 +289,7 @@
       document.querySelectorAll(".page-context__picker").forEach((other) => closePicker(other));
       menu.hidden = !opening;
       trigger.setAttribute("aria-expanded", String(opening));
+      if (opening) positionMenu(picker);
     });
     trigger.addEventListener("keydown", (event) => {
       if (event.key === "Escape") {
@@ -220,6 +301,7 @@
       event.preventDefault();
       menu.hidden = false;
       trigger.setAttribute("aria-expanded", "true");
+      positionMenu(picker);
       const options = [...menu.querySelectorAll("[role=option]")];
       (event.key === "ArrowDown" ? options[0] : options.at(-1))?.focus();
     });
@@ -255,7 +337,7 @@
         : available[0];
 
     buildControl(strip, available, initial, requested && !available.includes(requested) ? requested : null);
-    if (requested && available.includes(requested)) apply(requested);
+    apply(initial);
   };
 
   let prevTabState = null;
@@ -371,7 +453,10 @@
       setTimeout(initSearchModalDefaults, 150);
     }
   });
-  window.addEventListener("resize", syncTabIndicator);
+  window.addEventListener("resize", () => {
+    syncTabIndicator();
+    document.querySelectorAll(".page-context__picker").forEach(positionMenu);
+  });
 
   const DEFAULT_SEARCH_TOPICS = [
     {
diff --git a/docs/llms.txt b/docs/llms.txt
index 6ce23a5..9869569 100644
--- a/docs/llms.txt
+++ b/docs/llms.txt
@@ -32,19 +32,19 @@ Follow the Markdown links below to read canonical page content without cloning t
 - [Gemma 3n model card](https://gemma.ai/models/gemma-3n/model-card.md): Model card for Gemma 3n: on-device inputs and outputs, training data, evaluation results, and intended use.
 - [DiffusionGemma model overview](https://gemma.ai/models/diffusiongemma.md): DiffusionGemma is an experimental open model that explores text diffusion, an exceptionally fast approach to text generation.
 - [DiffusionGemma model card](https://gemma.ai/models/diffusiongemma/model-card.md): DiffusionGemma is a generative model built by Google DeepMind.
-- [Diffusion in Text Generation Explained](https://gemma.ai/models/diffusiongemma/explained.md): To understand DiffusionGemma, it helps to examine the core limitations of standard language models and how text-based diffusion differs.
-- [Generate output](https://gemma.ai/models/diffusiongemma/inference.md): Generate text with DiffusionGemma using Hugging Face Transformers, including the GPU memory the model requires.
+- [DiffusionGemma — Diffusion in Text Generation Explained](https://gemma.ai/models/diffusiongemma/explained.md): To understand DiffusionGemma, it helps to examine the core limitations of standard language models and how text-based diffusion differs.
+- [DiffusionGemma — Generate output](https://gemma.ai/models/diffusiongemma/inference.md): Generate text with DiffusionGemma using Hugging Face Transformers, including the GPU memory the model requires.
 - [FunctionGemma model overview](https://gemma.ai/models/functiongemma.md): FunctionGemma is a specialized version of our Gemma 3 270M model tuned for function calling.
 - [FunctionGemma model card](https://gemma.ai/models/functiongemma/model-card.md): Model card for FunctionGemma: function-calling inputs and outputs, training data, evaluation results, and intended use.
 - [FunctionGemma formatting and best practices](https://gemma.ai/models/functiongemma/formatting.md): FunctionGemma is a specialized version of the Gemma 3 270M model, trained specifically for function calling (i.e., tool use).
-- [Function calling with Hugging Face Transformers](https://gemma.ai/models/functiongemma/inference.md): Call functions with FunctionGemma in Hugging Face Transformers, passing tools as manual JSON schemas or as raw Python functions.
+- [FunctionGemma — Function calling with Hugging Face Transformers](https://gemma.ai/models/functiongemma/inference.md): Call functions with FunctionGemma in Hugging Face Transformers, passing tools as manual JSON schemas or as raw Python functions.
 - [Full function calling sequence with FunctionGemma](https://gemma.ai/models/functiongemma/full-sequence.md): The complete FunctionGemma tool-calling cycle: the model's call, the developer's execution, and the final response built from the tool's output.
 - [Fine-tuning with FunctionGemma](https://gemma.ai/models/functiongemma/fine-tuning.md): This guide demonstrates how to fine-tune FunctionGemma for tool calling.
 - [EmbeddingGemma model overview](https://gemma.ai/models/embeddinggemma.md): EmbeddingGemma 2 is a lightweight, open-source 740M parameter model designed for unified multimodal embeddings.
 - [EmbeddingGemma model card](https://gemma.ai/models/embeddinggemma/model-card-v1.md): Model card for EmbeddingGemma: embedding inputs and outputs, training data, evaluation results, and intended use.
 - [EmbeddingGemma 2 model card](https://gemma.ai/models/embeddinggemma/model-card-v2.md): Model card for EmbeddingGemma 2: architecture, multimodal capabilities, benchmarks, best practices, training data, and intended use.
-- [Generate Embeddings with Sentence Transformers](https://gemma.ai/models/embeddinggemma/inference.md): Generate text embeddings, compute semantic similarity, apply task prompts, and build RAG and zero-shot classifiers with EmbeddingGemma 2 and Sentence Transformers.
-- [Multimodal Embeddings with Sentence Transformers](https://gemma.ai/models/embeddinggemma/multimodal.md): Generate unified image, audio, video, and interleaved multimodal embeddings with EmbeddingGemma 2 and Sentence Transformers.
+- [EmbeddingGemma — Generate Embeddings with Sentence Transformers](https://gemma.ai/models/embeddinggemma/inference.md): Generate text embeddings, compute semantic similarity, apply task prompts, and build RAG and zero-shot classifiers with EmbeddingGemma 2 and Sentence Transformers.
+- [EmbeddingGemma — Multimodal Embeddings with Sentence Transformers](https://gemma.ai/models/embeddinggemma/multimodal.md): Generate unified image, audio, video, and interleaved multimodal embeddings with EmbeddingGemma 2 and Sentence Transformers.
 - [Fine-tune EmbeddingGemma with Sentence Transformers](https://gemma.ai/models/embeddinggemma/fine-tuning.md): Adapt EmbeddingGemma 2 to specialized domain terminology using contrastive triplet fine-tuning with Sentence Transformers.
 - [Fine-Tune Multimodal EmbeddingGemma 2 with Sentence Transformers](https://gemma.ai/models/embeddinggemma/fine-tuning-multimodal.md): Adapt EmbeddingGemma 2 across text and visual inputs using contrastive multimodal triplet fine-tuning with Sentence Transformers.
 - [PaliGemma](https://gemma.ai/models/paligemma.md): PaliGemma 2 and PaliGemma are lightweight open vision-language models (VLM) inspired by PaLI-3, and based on open components like the SigLIP vision model and the Gemma language model.
@@ -59,14 +59,14 @@ Follow the Markdown links below to read canonical page content without cloning t
 - [RecurrentGemma](https://gemma.ai/models/recurrentgemma.md): RecurrentGemma is an open model based on Griffin, a hybrid model architecture that mixes gated linear recurrences with local sliding window attention.
 - [RecurrentGemma model card](https://gemma.ai/models/recurrentgemma/model-card.md): RecurrentGemma is a family of open language models built on a novel recurrent architecture developed at Google.
 - [Inference with RecurrentGemma using JAX and Flax](https://gemma.ai/models/recurrentgemma/inference.md): Run sampling and inference with RecurrentGemma 2B Instruct using Google DeepMind's recurrentgemma library and JAX.
-- [Fine-tuning using JAX and Flax](https://gemma.ai/models/recurrentgemma/fine-tuning.md): Fine-tune RecurrentGemma 2B Instruct for English-French translation using JAX, Flax, and the MTNT noisy-text dataset.
+- [RecurrentGemma — Fine-tuning using JAX and Flax](https://gemma.ai/models/recurrentgemma/fine-tuning.md): Fine-tune RecurrentGemma 2B Instruct for English-French translation using JAX, Flax, and the MTNT noisy-text dataset.
 - [DataGemma](https://gemma.ai/models/datagemma.md): DataGemma is a research tool that lets users ask questions in plain language and receive answers based on publicly available statistical data in the Data Commons repository.
 - [Gemma Scope](https://gemma.ai/models/gemma-scope.md): Gemma Scope 2 is a comprehensive, open suite of interpretability tools designed for the Gemma 3 model collection.
 - [Gemma for abstractive proposition segmentation (APS)](https://gemma.ai/models/gemma-aps.md): Abstractive proposition segmentation (APS) is an analysis technique that breaks down text information into component parts.
 - [Releases](https://gemma.ai/models/releases.md): Dated release history for every Gemma model, from Gemma 1 through Gemma 4 and the specialized variants.
 
 ## Run
-- [Gemma 4 — Overview](https://gemma.ai/deployment.md): Choose an execution framework and a Gemma variant for your hardware, then run your first generation request.
+- [Overview](https://gemma.ai/deployment.md): Choose an execution framework and a Gemma variant for your hardware, then run your first generation request.
 - [Ollama](https://gemma.ai/deployment/ollama.md): Run Gemma 4 locally with Ollama — model tags, image input, and the OpenAI-compatible server.
 - [llama.cpp](https://gemma.ai/deployment/llamacpp.md): Run Gemma 4 with llama.cpp — GGUF repos for every size, image/audio/video input, and the OpenAI-compatible server.
 - [LM Studio](https://gemma.ai/deployment/lmstudio.md): Run Gemma 4 with LM Studio — desktop app, lms CLI, and the OpenAI-compatible local server.
diff --git a/docs/models/datagemma/index.md b/docs/models/datagemma/index.md
index 2a914cf..7b9c979 100644
--- a/docs/models/datagemma/index.md
+++ b/docs/models/datagemma/index.md
@@ -9,7 +9,7 @@ source_url: "https://ai.google.dev/gemma/docs/datagemma.md.txt"
 DataGemma is a research tool that lets users ask questions in plain language and receive
 answers based on publicly available statistical data in the
 [Data Commons](https://datacommons.org/) repository. The tool uses
-specially built versions of [Gemma](../gemma-4/index.md),
+specially built versions of [Gemma 2](../gemma-2/index.md),
 the [Gemini API](https://ai.google.dev/gemini-api) with Gemini 1.5 Pro, and a set of libraries
 specifically designed to work with Data Commons.
 
diff --git a/docs/models/diffusiongemma/model-card.md b/docs/models/diffusiongemma/model-card.md
index e7bb0bf..fed31be 100644
--- a/docs/models/diffusiongemma/model-card.md
+++ b/docs/models/diffusiongemma/model-card.md
@@ -12,7 +12,7 @@ source_url: "https://ai.google.dev/gemma/docs/diffusiongemma/model_card.md.txt"
 [Documentation](index.md)
 
 
-**License** : [Apache 2.0](https://ai.google.dev/gemma/apache_2) \| **Authors** : [Google DeepMind](https://deepmind.google/models/gemma/)
+**License** : [Apache 2.0](https://gemma.ai/resources/apache-2) \| **Authors** : [Google DeepMind](https://deepmind.google/models/gemma/)
 
 DiffusionGemma is a generative model built by Google DeepMind. Based on the 26B
 A4B Mixture-of-Experts (MoE) Gemma 4 architecture, DiffusionGemma generates
diff --git a/docs/models/embeddinggemma/fine-tuning-multimodal.md b/docs/models/embeddinggemma/fine-tuning-multimodal.md
index 026e64f..05ab8ce 100644
--- a/docs/models/embeddinggemma/fine-tuning-multimodal.md
+++ b/docs/models/embeddinggemma/fine-tuning-multimodal.md
@@ -1,7 +1,7 @@
 ---
 title: Fine-Tune Multimodal EmbeddingGemma 2 with Sentence Transformers
 description: Adapt EmbeddingGemma 2 across text and visual inputs using contrastive multimodal triplet fine-tuning with Sentence Transformers.
-models: [embeddinggemma]
+models: [embeddinggemma-2]
 frameworks: [sentence-transformers, transformers]
 colab: https://colab.research.google.com/github/google-gemma/docs/blob/main/docs/notebooks/embeddinggemma-fine-tuning-multimodal.ipynb
 source_url: "https://ai.google.dev/gemma/docs/embeddinggemma/fine-tuning-multimodal-embeddinggemma-with-sentence-transformers.md.txt"
diff --git a/docs/models/embeddinggemma/fine-tuning.md b/docs/models/embeddinggemma/fine-tuning.md
index 93428b5..5077709 100644
--- a/docs/models/embeddinggemma/fine-tuning.md
+++ b/docs/models/embeddinggemma/fine-tuning.md
@@ -1,7 +1,7 @@
 ---
 title: Fine-tune EmbeddingGemma with Sentence Transformers
 description: Adapt EmbeddingGemma 2 to specialized domain terminology using contrastive triplet fine-tuning with Sentence Transformers.
-models: [embeddinggemma]
+models: [embeddinggemma-2]
 frameworks: [sentence-transformers, transformers]
 colab: https://colab.research.google.com/github/google-gemma/docs/blob/main/docs/notebooks/embeddinggemma-fine-tuning.ipynb
 source_url: "https://ai.google.dev/gemma/docs/embeddinggemma/fine-tuning-embeddinggemma-with-sentence-transformers.md.txt"
diff --git a/docs/models/embeddinggemma/index.md b/docs/models/embeddinggemma/index.md
index a47af3d..5fcb391 100644
--- a/docs/models/embeddinggemma/index.md
+++ b/docs/models/embeddinggemma/index.md
@@ -1,7 +1,7 @@
 ---
 title: EmbeddingGemma model overview
 description: EmbeddingGemma 2 is a lightweight, open-source 740M parameter model designed for unified multimodal embeddings.
-models: [embeddinggemma]
+models: [embeddinggemma-2, embeddinggemma]
 source_url: "https://ai.google.dev/gemma/docs/embeddinggemma.md.txt"
 ---
 # EmbeddingGemma model overview {: #hidden-title style="display: none;" }
diff --git a/docs/models/embeddinggemma/inference.md b/docs/models/embeddinggemma/inference.md
index 95a3fbc..b744e15 100644
--- a/docs/models/embeddinggemma/inference.md
+++ b/docs/models/embeddinggemma/inference.md
@@ -1,7 +1,7 @@
 ---
 title: Generate Embeddings with Sentence Transformers
 description: Generate text embeddings, compute semantic similarity, apply task prompts, and build RAG and zero-shot classifiers with EmbeddingGemma 2 and Sentence Transformers.
-models: [embeddinggemma]
+models: [embeddinggemma-2]
 frameworks: [sentence-transformers, transformers]
 colab: https://colab.research.google.com/github/google-gemma/docs/blob/main/docs/notebooks/embeddinggemma-inference.ipynb
 source_url: "https://ai.google.dev/gemma/docs/embeddinggemma/inference-embeddinggemma-with-sentence-transformers.md.txt"
diff --git a/docs/models/embeddinggemma/model-card-v2.md b/docs/models/embeddinggemma/model-card-v2.md
index fdaf7a6..c159207 100644
--- a/docs/models/embeddinggemma/model-card-v2.md
+++ b/docs/models/embeddinggemma/model-card-v2.md
@@ -1,7 +1,7 @@
 ---
 title: EmbeddingGemma 2 model card
 description: "Model card for EmbeddingGemma 2: architecture, multimodal capabilities, benchmarks, best practices, training data, and intended use."
-models: [embeddinggemma]
+models: [embeddinggemma-2]
 source_url: "https://ai.google.dev/gemma/docs/embeddinggemma/model_card_2.md.txt"
 ---
 # EmbeddingGemma 2 model card
diff --git a/docs/models/embeddinggemma/multimodal.md b/docs/models/embeddinggemma/multimodal.md
index d8844d2..3418986 100644
--- a/docs/models/embeddinggemma/multimodal.md
+++ b/docs/models/embeddinggemma/multimodal.md
@@ -1,7 +1,7 @@
 ---
 title: Multimodal Embeddings with Sentence Transformers
 description: Generate unified image, audio, video, and interleaved multimodal embeddings with EmbeddingGemma 2 and Sentence Transformers.
-models: [embeddinggemma]
+models: [embeddinggemma-2]
 frameworks: [sentence-transformers, transformers]
 colab: https://colab.research.google.com/github/google-gemma/docs/blob/main/docs/notebooks/embeddinggemma-multimodal.ipynb
 source_url: "https://ai.google.dev/gemma/docs/embeddinggemma/multimodal-embeddinggemma-with-sentence-transformers.md.txt"
diff --git a/docs/models/functiongemma/fine-tuning.md b/docs/models/functiongemma/fine-tuning.md
index ccb1d20..34f7aa0 100644
--- a/docs/models/functiongemma/fine-tuning.md
+++ b/docs/models/functiongemma/fine-tuning.md
@@ -608,7 +608,7 @@ Now that you have a fine-tuned model, consider the following steps to move towar
   <div class="card-title">Fine-tune Gemma</div>
   <p class="card-desc">QLoRA, full fine-tunes and distributed training.</p>
 </a>
-<a class="card" href="https://github.com/google-gemma/cookbook/blob/main/FunctionGemma/%5BFunctionGemma%5DFinetune_FunctionGemma_270M_for_Mobile_Actions_with_Hugging_Face.ipynb" target="_blank" rel="noopener" data-model="functiongemma">
+<a class="card" href="https://github.com/google-gemma/cookbook/tree/main/.archive/FunctionGemma" target="_blank" rel="noopener" data-model="functiongemma">
   <div class="card-title">Mobile Actions notebook <span class="card-ext">↗</span></div>
   <p class="card-desc">Fine-tune FunctionGemma 270M for on-device actions.</p>
 </a>
diff --git a/docs/models/functiongemma/model-card.md b/docs/models/functiongemma/model-card.md
index e6d2281..63043b0 100644
--- a/docs/models/functiongemma/model-card.md
+++ b/docs/models/functiongemma/model-card.md
@@ -128,7 +128,7 @@ capable of generalizing across multiple tasks. This is specially suitable for
 foundation models, including large language models like these ones.  
 
 Together, JAX and ML Pathways are used as described in the [paper about the
-Gemini family of models](https://goo.gle/gemma2report); *"the 'single
+Gemini family of models](https://arxiv.org/abs/2312.11805); *"the 'single
 controller' programming model of Jax and Pathways allows a single Python process
 to orchestrate the entire training run, dramatically simplifying the development
 workflow."*
diff --git a/docs/models/gemma-2/index.md b/docs/models/gemma-2/index.md
index bcd7da6..3c984a6 100644
--- a/docs/models/gemma-2/index.md
+++ b/docs/models/gemma-2/index.md
@@ -111,7 +111,7 @@ capable of generalizing across multiple tasks. This is specially suitable for
 foundation models, including large language models like these ones.
 
 Together, JAX and ML Pathways are used as described in the
-[paper about the Gemini family of models](https://goo.gle/gemma2report){ data-model="gemma-2" }; "the 'single
+[paper about the Gemini family of models](https://arxiv.org/abs/2312.11805); "the 'single
 controller' programming model of Jax and Pathways allows a single Python
 process to orchestrate the entire training run, dramatically simplifying the
 development workflow."
diff --git a/docs/models/gemma-3n/model-card.md b/docs/models/gemma-3n/model-card.md
index 912d464..2c45758 100644
--- a/docs/models/gemma-3n/model-card.md
+++ b/docs/models/gemma-3n/model-card.md
@@ -124,7 +124,7 @@ capable of generalizing across multiple tasks. This is specially suitable for
 foundation models, including large language models like these ones.
 
 Together, JAX and ML Pathways are used as described in the
-[paper about the Gemini family of models](https://goo.gle/gemma2report){ data-model="gemma-2" }:
+[paper about the Gemini family of models](https://arxiv.org/abs/2312.11805):
 *"the 'single controller' programming model of Jax and Pathways allows a single
 Python process to orchestrate the entire training run, dramatically simplifying
 the development workflow."*
diff --git a/docs/models/gemma-4/index.md b/docs/models/gemma-4/index.md
index f24d342..b28757b 100644
--- a/docs/models/gemma-4/index.md
+++ b/docs/models/gemma-4/index.md
@@ -17,7 +17,7 @@ Gemma is a family of generative artificial intelligence models and you can
 use them in a wide variety of generation tasks, including question answering,
 summarization, and reasoning. Gemma models are provided with open weights and
 permit responsible
-[commercial use][gemma-terms],
+[commercial use](../../resources/apache-2.md),
 allowing you to tune and deploy them in your own projects and applications.
 
 Gemma 4 model family spans four distinct architectures tailored for specific
diff --git a/docs/models/gemma-4/model-card.md b/docs/models/gemma-4/model-card.md
index 5b36c73..42064f1 100644
--- a/docs/models/gemma-4/model-card.md
+++ b/docs/models/gemma-4/model-card.md
@@ -14,7 +14,7 @@ source_url: "https://ai.google.dev/gemma/docs/core/model_card_4.md.txt"
 [Documentation](index.md) \|
 [Technical Report](https://arxiv.org/abs/2607.02770)
 
-**License** : [Apache 2.0](https://ai.google.dev/gemma/apache_2) \| **Authors** : [Google DeepMind](https://deepmind.google/models/gemma/)
+**License** : [Apache 2.0](https://gemma.ai/resources/apache-2) \| **Authors** : [Google DeepMind](https://deepmind.google/models/gemma/)
 
 </div>
 
diff --git a/docs/models/gemma-scope/index.md b/docs/models/gemma-scope/index.md
index 7c755a7..6b0b3db 100644
--- a/docs/models/gemma-scope/index.md
+++ b/docs/models/gemma-scope/index.md
@@ -38,7 +38,7 @@ Read about the new architecture, training data, and capabilities of Gemma Scope
 
 Access the weights, code, and documentation for the Gemma 3 interpretability suite.
 
-### [Run in Colab](https://colab.sandbox.google.com/drive/1NhWjg7n0nhfW--CjtsOdw5A5J_-Bzn4r?usp=sharing)
+### [Run in Colab](https://colab.research.google.com/drive/1NhWjg7n0nhfW--CjtsOdw5A5J_-Bzn4r?usp=sharing)
 
 Try the interactive tutorial to visualize features and modify model behavior.
 
diff --git a/docs/models/paligemma/fine-tuning.md b/docs/models/paligemma/fine-tuning.md
index e9ab5c3..3ebd446 100644
--- a/docs/models/paligemma/fine-tuning.md
+++ b/docs/models/paligemma/fine-tuning.md
@@ -1,7 +1,7 @@
 ---
 title: Fine-tune PaliGemma with JAX and Flax
 description: This notebook shows how to fine-tune PaliGemma on a vision-language task with JAX.
-models: [paligemma]
+models: [paligemma-2, paligemma]
 frameworks: [jax]
 colab: https://colab.research.google.com/github/google-gemma/docs/blob/main/docs/notebooks/paligemma-fine-tuning.ipynb
 source_url: "https://ai.google.dev/gemma/docs/paligemma/fine-tuning-paligemma.md.txt"
diff --git a/docs/models/paligemma/index.md b/docs/models/paligemma/index.md
index f89ced5..d558c79 100644
--- a/docs/models/paligemma/index.md
+++ b/docs/models/paligemma/index.md
@@ -1,7 +1,7 @@
 ---
 title: PaliGemma
 description: PaliGemma 2 and PaliGemma are lightweight open vision-language models (VLM) inspired by PaLI-3, and based on open components like the SigLIP vision model and the Gemma language model.
-models: [paligemma]
+models: [paligemma-2, paligemma]
 source_url: "https://ai.google.dev/gemma/docs/paligemma.md.txt"
 ---
 # PaliGemma
diff --git a/docs/models/paligemma/inference.md b/docs/models/paligemma/inference.md
index cd81289..dd1d11f 100644
--- a/docs/models/paligemma/inference.md
+++ b/docs/models/paligemma/inference.md
@@ -1,7 +1,7 @@
 ---
 title: Generate PaliGemma output with Keras
 description: PaliGemma models have multimodal capabilities, allowing you to generate output using both text and image input data.
-models: [paligemma]
+models: [paligemma-2]
 frameworks: [keras]
 colab: https://colab.research.google.com/github/google-gemma/docs/blob/main/docs/notebooks/paligemma-inference.ipynb
 source_url: "https://ai.google.dev/gemma/docs/paligemma/inference-with-keras.md.txt"
diff --git a/docs/models/paligemma/model-card-v2.md b/docs/models/paligemma/model-card-v2.md
index dc69621..e5c6e9a 100644
--- a/docs/models/paligemma/model-card-v2.md
+++ b/docs/models/paligemma/model-card-v2.md
@@ -1,7 +1,7 @@
 ---
 title: PaliGemma 2 model card
 description: PaliGemma 2 is an update of the PaliGemma vision-language model (VLM) which incorporates the capabilities of the Gemma 2 models.
-models: [paligemma]
+models: [paligemma-2]
 source_url: "https://ai.google.dev/gemma/docs/paligemma/model-card-2.md.txt"
 ---
 # Model card
@@ -38,7 +38,7 @@ PaliGemma 2 is the composition of a
 [Transformer decoder](https://arxiv.org/abs/1706.03762) and a
 [Vision Transformer image encoder](https://arxiv.org/abs/2010.11929).
 The text decoder is initialized from
-[Gemma 2](../gemma-4/index.md) in the 2B, 9B, and 27B
+[Gemma 2](../gemma-2/index.md) in the 2B, 9B, and 27B
 parameter sizes. The image encoder is initialized from
 [SigLIP-So400m/14](https://colab.research.google.com/github/google-research/big_vision/blob/main/big_vision/configs/proj/image_text/SigLIP_demo.ipynb).
 Similar to the original PaliGemma model, PaliGemma 2 is trained following the
diff --git a/docs/models/paligemma/prompting.md b/docs/models/paligemma/prompting.md
index 7d29d07..dc550bf 100644
--- a/docs/models/paligemma/prompting.md
+++ b/docs/models/paligemma/prompting.md
@@ -1,7 +1,7 @@
 ---
 title: PaliGemma prompt and system instructions
 description: This page describes prompt formatting and system instructions for PaliGemma models.
-models: [paligemma]
+models: [paligemma-2, paligemma]
 source_url: "https://ai.google.dev/gemma/docs/paligemma/prompt-system-instructions.md.txt"
 ---
 # PaliGemma prompt and system instructions
diff --git a/docs/models/shieldgemma/index.md b/docs/models/shieldgemma/index.md
index e4c668a..1fcc3ed 100644
--- a/docs/models/shieldgemma/index.md
+++ b/docs/models/shieldgemma/index.md
@@ -1,7 +1,7 @@
 ---
 title: ShieldGemma
 description: ShieldGemma is a set of instruction tuned models for evaluating the safety of text and images against a set of defined safety policies.
-models: [shieldgemma]
+models: [shieldgemma-2, shieldgemma]
 source_url: "https://ai.google.dev/gemma/docs/shieldgemma.md.txt"
 ---
 # ShieldGemma {: #hidden-title style="display: none;" }
@@ -21,7 +21,7 @@ for your specific use case.
 
 
 ShieldGemma 2 is a 4B parameter model built to label images for safety.
-ShieldGemma 1 is built on [Gemma 2](../gemma-4/index.md) in 2B, 9B, and 27B parameter sizes.
+ShieldGemma 1 is built on [Gemma 2](../gemma-2/index.md) in 2B, 9B, and 27B parameter sizes.
 
 
 - **Content safety evaluation** — Evaluate the safety of prompt input and output responses against a set of defined safety policies.
diff --git a/docs/models/shieldgemma/model-card-v2.md b/docs/models/shieldgemma/model-card-v2.md
index 0fb80ca..4de8bf7 100644
--- a/docs/models/shieldgemma/model-card-v2.md
+++ b/docs/models/shieldgemma/model-card-v2.md
@@ -1,7 +1,7 @@
 ---
 title: ShieldGemma 2 model card
 description: ShieldGemma 2 is a model trained on Gemma 3's 4B IT checkpoint for image safety classification across key categories that takes in images and outputs safety labels per policy.
-models: [shieldgemma]
+models: [shieldgemma-2]
 source_url: "https://ai.google.dev/gemma/docs/shieldgemma/model_card_2.md.txt"
 ---
 # ShieldGemma model card
@@ -11,8 +11,8 @@ source_url: "https://ai.google.dev/gemma/docs/shieldgemma/model_card_2.md.txt"
 **Resources and Technical Documentation**:
 
 - [Responsible Generative AI Toolkit][responsible-ai]
-- [ShieldGemma on Kaggle](https://www.kaggle.com/models/google/shieldgemma){ data-model="shieldgemma" }
-- [ShieldGemma on Hugging Face Hub](https://huggingface.co/models?search=shieldgemma){ data-model="shieldgemma" }
+- [ShieldGemma on Kaggle](https://www.kaggle.com/models/google/shieldgemma-2){ data-model="shieldgemma-2" }
+- [ShieldGemma on Hugging Face Hub](https://huggingface.co/models?search=shieldgemma){ data-model="shieldgemma-2" }
 
 **Terms of Use** : [Terms][gemma-terms]
 
diff --git a/docs/notebooks/audio-qlora-gemma-4-12b.ipynb b/docs/notebooks/audio-qlora-gemma-4-12b.ipynb
index 6f558c6..7ef1596 100644
--- a/docs/notebooks/audio-qlora-gemma-4-12b.ipynb
+++ b/docs/notebooks/audio-qlora-gemma-4-12b.ipynb
@@ -302,8 +302,8 @@
     "\n",
     "To ensure the model learns strictly to transcribe the audio rather than memorizing prompt tokens:\n",
     "\n",
-    "- The following code parses <code><span class=\"token-turn\"><|turn></span><span class=\"token-role\">model</span>\\n</code> in each sequence.\n",
-    "- All tokens prior to <code><span class=\"token-turn\"><|turn></span><span class=\"token-role\">model</span>\\n</code>  (including audio prefix markers and user instruction) are masked to `-100`.\n",
+    "- The following code parses <code><span class=\"token-turn\">&lt;|turn&gt;</span><span class=\"token-role\">model</span>\\n</code> in each sequence.\n",
+    "- All tokens prior to <code><span class=\"token-turn\">&lt;|turn&gt;</span><span class=\"token-role\">model</span>\\n</code>  (including audio prefix markers and user instruction) are masked to `-100`.\n",
     "- Only assistant response tokens contribute to the cross-entropy loss."
    ]
   },
diff --git a/docs/notebooks/audio-qlora-gemma-4-e2b.ipynb b/docs/notebooks/audio-qlora-gemma-4-e2b.ipynb
index dee25a9..fdb3675 100644
--- a/docs/notebooks/audio-qlora-gemma-4-e2b.ipynb
+++ b/docs/notebooks/audio-qlora-gemma-4-e2b.ipynb
@@ -302,8 +302,8 @@
     "\n",
     "To ensure the model learns strictly to transcribe the audio rather than memorizing prompt tokens:\n",
     "\n",
-    "- The following code parses <code><span class=\"token-turn\"><|turn></span><span class=\"token-role\">model</span>\\n</code> in each sequence.\n",
-    "- All tokens prior to <code><span class=\"token-turn\"><|turn></span><span class=\"token-role\">model</span>\\n</code>  (including audio prefix markers and user instruction) are masked to `-100`.\n",
+    "- The following code parses <code><span class=\"token-turn\">&lt;|turn&gt;</span><span class=\"token-role\">model</span>\\n</code> in each sequence.\n",
+    "- All tokens prior to <code><span class=\"token-turn\">&lt;|turn&gt;</span><span class=\"token-role\">model</span>\\n</code>  (including audio prefix markers and user instruction) are masked to `-100`.\n",
     "- Only assistant response tokens contribute to the cross-entropy loss."
    ]
   },
diff --git a/docs/notebooks/audio-qlora-gemma-4-e4b.ipynb b/docs/notebooks/audio-qlora-gemma-4-e4b.ipynb
index a322d82..7a51e66 100644
--- a/docs/notebooks/audio-qlora-gemma-4-e4b.ipynb
+++ b/docs/notebooks/audio-qlora-gemma-4-e4b.ipynb
@@ -302,8 +302,8 @@
     "\n",
     "To ensure the model learns strictly to transcribe the audio rather than memorizing prompt tokens:\n",
     "\n",
-    "- The following code parses <code><span class=\"token-turn\"><|turn></span><span class=\"token-role\">model</span>\\n</code> in each sequence.\n",
-    "- All tokens prior to <code><span class=\"token-turn\"><|turn></span><span class=\"token-role\">model</span>\\n</code>  (including audio prefix markers and user instruction) are masked to `-100`.\n",
+    "- The following code parses <code><span class=\"token-turn\">&lt;|turn&gt;</span><span class=\"token-role\">model</span>\\n</code> in each sequence.\n",
+    "- All tokens prior to <code><span class=\"token-turn\">&lt;|turn&gt;</span><span class=\"token-role\">model</span>\\n</code>  (including audio prefix markers and user instruction) are masked to `-100`.\n",
     "- Only assistant response tokens contribute to the cross-entropy loss."
    ]
   },
diff --git a/docs/notebooks/function-calling-gemma-4-12b.ipynb b/docs/notebooks/function-calling-gemma-4-12b.ipynb
index eb8e70a..f66acea 100644
--- a/docs/notebooks/function-calling-gemma-4-12b.ipynb
+++ b/docs/notebooks/function-calling-gemma-4-12b.ipynb
@@ -311,40 +311,20 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "> **Note:** For optimal results, append the tool execution result to your message history using the specific format below. This ensures the chat template correctly generates the required token structure (e.g., `response:get_current_weather{temperature:15,weather:<|\"|>sunny<|\"|>}`)."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "language": "python"
-   },
-   "outputs": [],
-   "source": [
+    "> **Note:** For optimal results, append the tool execution result to your message history using the specific format below. This ensures the chat template correctly generates the required token structure (e.g., `response:get_current_weather{temperature:15,weather:<|\"|>sunny<|\"|>}`).\n",
+    "\n",
+    "```python\n",
     "\"tool_responses\": [\n",
     "  {\n",
     "    \"name\": function_name,\n",
     "    \"response\": function_response\n",
     "  }\n",
-    "]\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "In case of multiple independent requests:"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "language": "python"
-   },
-   "outputs": [],
-   "source": [
+    "]\n",
+    "```\n",
+    "\n",
+    "In case of multiple independent requests:\n",
+    "\n",
+    "```python\n",
     "\"tool_responses\": [\n",
     "  {\n",
     "    \"name\": function_name_1,\n",
@@ -354,13 +334,9 @@
     "    \"name\": function_name_2,\n",
     "    \"response\": function_response_2\n",
     "  }\n",
-    "]\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
+    "]\n",
+    "```\n",
+    "\n",
     "### Final Response\n",
     "\n",
     "Finally, Gemma reads the tool response and reply to the user."
diff --git a/docs/notebooks/function-calling-gemma-4-26b-a4b.ipynb b/docs/notebooks/function-calling-gemma-4-26b-a4b.ipynb
index d058323..031c575 100644
--- a/docs/notebooks/function-calling-gemma-4-26b-a4b.ipynb
+++ b/docs/notebooks/function-calling-gemma-4-26b-a4b.ipynb
@@ -311,40 +311,20 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "> **Note:** For optimal results, append the tool execution result to your message history using the specific format below. This ensures the chat template correctly generates the required token structure (e.g., `response:get_current_weather{temperature:15,weather:<|\"|>sunny<|\"|>}`)."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "language": "python"
-   },
-   "outputs": [],
-   "source": [
+    "> **Note:** For optimal results, append the tool execution result to your message history using the specific format below. This ensures the chat template correctly generates the required token structure (e.g., `response:get_current_weather{temperature:15,weather:<|\"|>sunny<|\"|>}`).\n",
+    "\n",
+    "```python\n",
     "\"tool_responses\": [\n",
     "  {\n",
     "    \"name\": function_name,\n",
     "    \"response\": function_response\n",
     "  }\n",
-    "]\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "In case of multiple independent requests:"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "language": "python"
-   },
-   "outputs": [],
-   "source": [
+    "]\n",
+    "```\n",
+    "\n",
+    "In case of multiple independent requests:\n",
+    "\n",
+    "```python\n",
     "\"tool_responses\": [\n",
     "  {\n",
     "    \"name\": function_name_1,\n",
@@ -354,13 +334,9 @@
     "    \"name\": function_name_2,\n",
     "    \"response\": function_response_2\n",
     "  }\n",
-    "]\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
+    "]\n",
+    "```\n",
+    "\n",
     "### Final Response\n",
     "\n",
     "Finally, Gemma reads the tool response and reply to the user."
diff --git a/docs/notebooks/function-calling-gemma-4-31b.ipynb b/docs/notebooks/function-calling-gemma-4-31b.ipynb
index 73b8537..7308266 100644
--- a/docs/notebooks/function-calling-gemma-4-31b.ipynb
+++ b/docs/notebooks/function-calling-gemma-4-31b.ipynb
@@ -311,40 +311,20 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "> **Note:** For optimal results, append the tool execution result to your message history using the specific format below. This ensures the chat template correctly generates the required token structure (e.g., `response:get_current_weather{temperature:15,weather:<|\"|>sunny<|\"|>}`)."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "language": "python"
-   },
-   "outputs": [],
-   "source": [
+    "> **Note:** For optimal results, append the tool execution result to your message history using the specific format below. This ensures the chat template correctly generates the required token structure (e.g., `response:get_current_weather{temperature:15,weather:<|\"|>sunny<|\"|>}`).\n",
+    "\n",
+    "```python\n",
     "\"tool_responses\": [\n",
     "  {\n",
     "    \"name\": function_name,\n",
     "    \"response\": function_response\n",
     "  }\n",
-    "]\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "In case of multiple independent requests:"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "language": "python"
-   },
-   "outputs": [],
-   "source": [
+    "]\n",
+    "```\n",
+    "\n",
+    "In case of multiple independent requests:\n",
+    "\n",
+    "```python\n",
     "\"tool_responses\": [\n",
     "  {\n",
     "    \"name\": function_name_1,\n",
@@ -354,13 +334,9 @@
     "    \"name\": function_name_2,\n",
     "    \"response\": function_response_2\n",
     "  }\n",
-    "]\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
+    "]\n",
+    "```\n",
+    "\n",
     "### Final Response\n",
     "\n",
     "Finally, Gemma reads the tool response and reply to the user."
diff --git a/docs/notebooks/function-calling-gemma-4-e2b.ipynb b/docs/notebooks/function-calling-gemma-4-e2b.ipynb
index 2995aec..8bf3348 100644
--- a/docs/notebooks/function-calling-gemma-4-e2b.ipynb
+++ b/docs/notebooks/function-calling-gemma-4-e2b.ipynb
@@ -311,40 +311,20 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "> **Note:** For optimal results, append the tool execution result to your message history using the specific format below. This ensures the chat template correctly generates the required token structure (e.g., `response:get_current_weather{temperature:15,weather:<|\"|>sunny<|\"|>}`)."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "language": "python"
-   },
-   "outputs": [],
-   "source": [
+    "> **Note:** For optimal results, append the tool execution result to your message history using the specific format below. This ensures the chat template correctly generates the required token structure (e.g., `response:get_current_weather{temperature:15,weather:<|\"|>sunny<|\"|>}`).\n",
+    "\n",
+    "```python\n",
     "\"tool_responses\": [\n",
     "  {\n",
     "    \"name\": function_name,\n",
     "    \"response\": function_response\n",
     "  }\n",
-    "]\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "In case of multiple independent requests:"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "language": "python"
-   },
-   "outputs": [],
-   "source": [
+    "]\n",
+    "```\n",
+    "\n",
+    "In case of multiple independent requests:\n",
+    "\n",
+    "```python\n",
     "\"tool_responses\": [\n",
     "  {\n",
     "    \"name\": function_name_1,\n",
@@ -354,13 +334,9 @@
     "    \"name\": function_name_2,\n",
     "    \"response\": function_response_2\n",
     "  }\n",
-    "]\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
+    "]\n",
+    "```\n",
+    "\n",
     "### Final Response\n",
     "\n",
     "Finally, Gemma reads the tool response and reply to the user."
diff --git a/docs/notebooks/function-calling-gemma-4-e4b.ipynb b/docs/notebooks/function-calling-gemma-4-e4b.ipynb
index efe83e0..309ecdf 100644
--- a/docs/notebooks/function-calling-gemma-4-e4b.ipynb
+++ b/docs/notebooks/function-calling-gemma-4-e4b.ipynb
@@ -311,40 +311,20 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "> **Note:** For optimal results, append the tool execution result to your message history using the specific format below. This ensures the chat template correctly generates the required token structure (e.g., `response:get_current_weather{temperature:15,weather:<|\"|>sunny<|\"|>}`)."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "language": "python"
-   },
-   "outputs": [],
-   "source": [
+    "> **Note:** For optimal results, append the tool execution result to your message history using the specific format below. This ensures the chat template correctly generates the required token structure (e.g., `response:get_current_weather{temperature:15,weather:<|\"|>sunny<|\"|>}`).\n",
+    "\n",
+    "```python\n",
     "\"tool_responses\": [\n",
     "  {\n",
     "    \"name\": function_name,\n",
     "    \"response\": function_response\n",
     "  }\n",
-    "]\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "In case of multiple independent requests:"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "language": "python"
-   },
-   "outputs": [],
-   "source": [
+    "]\n",
+    "```\n",
+    "\n",
+    "In case of multiple independent requests:\n",
+    "\n",
+    "```python\n",
     "\"tool_responses\": [\n",
     "  {\n",
     "    \"name\": function_name_1,\n",
@@ -354,13 +334,9 @@
     "    \"name\": function_name_2,\n",
     "    \"response\": function_response_2\n",
     "  }\n",
-    "]\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
+    "]\n",
+    "```\n",
+    "\n",
     "### Final Response\n",
     "\n",
     "Finally, Gemma reads the tool response and reply to the user."
diff --git a/docs/notebooks/run-gemma-with-hf-gemma-4-12b.ipynb b/docs/notebooks/run-gemma-with-hf-gemma-4-12b.ipynb
index c42d8f0..d44474c 100644
--- a/docs/notebooks/run-gemma-with-hf-gemma-4-12b.ipynb
+++ b/docs/notebooks/run-gemma-with-hf-gemma-4-12b.ipynb
@@ -37,7 +37,7 @@
     "\n",
     "Generating text, summarizing, and analysing content are just some of the tasks you can accomplish with Gemma open models. This tutorial shows you how to get started running Gemma using Hugging Face Transformers using both text and image input to generate text content. The Transformers Python library provides a API for accessing pre-trained generative AI models, including Gemma. For more information, see the [Transformers](https://huggingface.co/docs/transformers/en/index) documentation.\n",
     "\n",
-    "> **Note:** Gemma 3 and later models support input of both images and text. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../get-started/index.md).\n",
+    "> **Note:** Gemma 3 and later models support input of both images and text. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../models/paligemma/index.md).\n",
     "\n",
     "\n",
     "### Install Python packages\n",
diff --git a/docs/notebooks/run-gemma-with-hf-gemma-4-26b-a4b.ipynb b/docs/notebooks/run-gemma-with-hf-gemma-4-26b-a4b.ipynb
index 6c9aa38..72d6bdf 100644
--- a/docs/notebooks/run-gemma-with-hf-gemma-4-26b-a4b.ipynb
+++ b/docs/notebooks/run-gemma-with-hf-gemma-4-26b-a4b.ipynb
@@ -37,7 +37,7 @@
     "\n",
     "Generating text, summarizing, and analysing content are just some of the tasks you can accomplish with Gemma open models. This tutorial shows you how to get started running Gemma using Hugging Face Transformers using both text and image input to generate text content. The Transformers Python library provides a API for accessing pre-trained generative AI models, including Gemma. For more information, see the [Transformers](https://huggingface.co/docs/transformers/en/index) documentation.\n",
     "\n",
-    "> **Note:** Gemma 3 and later models support input of both images and text. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../get-started/index.md).\n",
+    "> **Note:** Gemma 3 and later models support input of both images and text. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../models/paligemma/index.md).\n",
     "\n",
     "\n",
     "### Install Python packages\n",
@@ -203,47 +203,6 @@
     "You can include multiple images in your prompt by including additional `\"type\": \"image\",` entries in the `content` list.\n",
     "> **Note:** Do not use `<|image|>`, `<start_of_image>` or `<image_soft_token>` tokens in text portion of a prompt template as this approach creates redundant tokens and processing errors.\n",
     "\n",
-    "## Generate text from audio data\n",
-    "\n",
-    "With [Gemma 4](../models/gemma-4/index.md) and [Gemma 3n](../models/gemma-3n/index.md), you can use audio data as part of your prompt. This section shows how to use the Transformers library to load and configure a Gemma model to use audio data and text input to generate text output.\n",
-    "\n",
-    "### Use a prompt template\n",
-    "\n",
-    "When generating content with audio, use a prompt template to structure your request. A prompt template allows you to specify input from specific roles, such as `user` or `model`, and is a required format for managing multi-turn chat interactions with Gemma models. The following example code shows how to constuct a prompt template for Gemma with audio data input:"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "language": "python"
-   },
-   "outputs": [],
-   "source": [
-    "from transformers import GenerationConfig\n",
-    "config = GenerationConfig.from_pretrained(MODEL_ID)\n",
-    "config.max_new_tokens = 512\n",
-    "gen_kwargs = dict(generation_config=config)\n",
-    "\n",
-    "messages = [\n",
-    "    {\n",
-    "        \"role\": \"user\",\n",
-    "        \"content\": [\n",
-    "            {\"type\": \"text\", \"text\": \"Transcribe the following speech segment in its original language. Follow these specific instructions for formatting the answer:\\n* Only output the transcription, with no newlines.\\n* When transcribing numbers, write the digits, i.e. write 1.7 and not one point seven, and write 3 instead of three.\"},\n",
-    "            {\"type\": \"audio\", \"audio\": \"https://ai.google.dev/gemma/docs/audio/roses-are.wav\"},\n",
-    "        ]\n",
-    "    }\n",
-    "]\n",
-    "\n",
-    "pipe(text=messages, return_full_text=False, generate_kwargs=gen_kwargs)\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "You can include multiple audio files in your prompt by including additional `\"type\": \"audio\",` entries in the `content` list.\n",
-    "> **Note:** Do not use `<|audio|>` or `<audio_soft_token>` tokens in text portion of a prompt template as this approach creates redundant tokens and processing errors.\n",
     "\n",
     "## Next steps\n",
     "\n",
diff --git a/docs/notebooks/run-gemma-with-hf-gemma-4-31b.ipynb b/docs/notebooks/run-gemma-with-hf-gemma-4-31b.ipynb
index 625ef6f..16c7d18 100644
--- a/docs/notebooks/run-gemma-with-hf-gemma-4-31b.ipynb
+++ b/docs/notebooks/run-gemma-with-hf-gemma-4-31b.ipynb
@@ -37,7 +37,7 @@
     "\n",
     "Generating text, summarizing, and analysing content are just some of the tasks you can accomplish with Gemma open models. This tutorial shows you how to get started running Gemma using Hugging Face Transformers using both text and image input to generate text content. The Transformers Python library provides a API for accessing pre-trained generative AI models, including Gemma. For more information, see the [Transformers](https://huggingface.co/docs/transformers/en/index) documentation.\n",
     "\n",
-    "> **Note:** Gemma 3 and later models support input of both images and text. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../get-started/index.md).\n",
+    "> **Note:** Gemma 3 and later models support input of both images and text. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../models/paligemma/index.md).\n",
     "\n",
     "\n",
     "### Install Python packages\n",
@@ -203,47 +203,6 @@
     "You can include multiple images in your prompt by including additional `\"type\": \"image\",` entries in the `content` list.\n",
     "> **Note:** Do not use `<|image|>`, `<start_of_image>` or `<image_soft_token>` tokens in text portion of a prompt template as this approach creates redundant tokens and processing errors.\n",
     "\n",
-    "## Generate text from audio data\n",
-    "\n",
-    "With [Gemma 4](../models/gemma-4/index.md) and [Gemma 3n](../models/gemma-3n/index.md), you can use audio data as part of your prompt. This section shows how to use the Transformers library to load and configure a Gemma model to use audio data and text input to generate text output.\n",
-    "\n",
-    "### Use a prompt template\n",
-    "\n",
-    "When generating content with audio, use a prompt template to structure your request. A prompt template allows you to specify input from specific roles, such as `user` or `model`, and is a required format for managing multi-turn chat interactions with Gemma models. The following example code shows how to constuct a prompt template for Gemma with audio data input:"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "language": "python"
-   },
-   "outputs": [],
-   "source": [
-    "from transformers import GenerationConfig\n",
-    "config = GenerationConfig.from_pretrained(MODEL_ID)\n",
-    "config.max_new_tokens = 512\n",
-    "gen_kwargs = dict(generation_config=config)\n",
-    "\n",
-    "messages = [\n",
-    "    {\n",
-    "        \"role\": \"user\",\n",
-    "        \"content\": [\n",
-    "            {\"type\": \"text\", \"text\": \"Transcribe the following speech segment in its original language. Follow these specific instructions for formatting the answer:\\n* Only output the transcription, with no newlines.\\n* When transcribing numbers, write the digits, i.e. write 1.7 and not one point seven, and write 3 instead of three.\"},\n",
-    "            {\"type\": \"audio\", \"audio\": \"https://ai.google.dev/gemma/docs/audio/roses-are.wav\"},\n",
-    "        ]\n",
-    "    }\n",
-    "]\n",
-    "\n",
-    "pipe(text=messages, return_full_text=False, generate_kwargs=gen_kwargs)\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "You can include multiple audio files in your prompt by including additional `\"type\": \"audio\",` entries in the `content` list.\n",
-    "> **Note:** Do not use `<|audio|>` or `<audio_soft_token>` tokens in text portion of a prompt template as this approach creates redundant tokens and processing errors.\n",
     "\n",
     "## Next steps\n",
     "\n",
diff --git a/docs/notebooks/run-gemma-with-hf-gemma-4-e2b.ipynb b/docs/notebooks/run-gemma-with-hf-gemma-4-e2b.ipynb
index d66809a..e911f28 100644
--- a/docs/notebooks/run-gemma-with-hf-gemma-4-e2b.ipynb
+++ b/docs/notebooks/run-gemma-with-hf-gemma-4-e2b.ipynb
@@ -37,7 +37,7 @@
     "\n",
     "Generating text, summarizing, and analysing content are just some of the tasks you can accomplish with Gemma open models. This tutorial shows you how to get started running Gemma using Hugging Face Transformers using both text and image input to generate text content. The Transformers Python library provides a API for accessing pre-trained generative AI models, including Gemma. For more information, see the [Transformers](https://huggingface.co/docs/transformers/en/index) documentation.\n",
     "\n",
-    "> **Note:** Gemma 3 and later models support input of both images and text. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../get-started/index.md).\n",
+    "> **Note:** Gemma 3 and later models support input of both images and text. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../models/paligemma/index.md).\n",
     "\n",
     "\n",
     "### Install Python packages\n",
diff --git a/docs/notebooks/run-gemma-with-hf-gemma-4-e4b.ipynb b/docs/notebooks/run-gemma-with-hf-gemma-4-e4b.ipynb
index 5cc92e6..6ec1bcd 100644
--- a/docs/notebooks/run-gemma-with-hf-gemma-4-e4b.ipynb
+++ b/docs/notebooks/run-gemma-with-hf-gemma-4-e4b.ipynb
@@ -37,7 +37,7 @@
     "\n",
     "Generating text, summarizing, and analysing content are just some of the tasks you can accomplish with Gemma open models. This tutorial shows you how to get started running Gemma using Hugging Face Transformers using both text and image input to generate text content. The Transformers Python library provides a API for accessing pre-trained generative AI models, including Gemma. For more information, see the [Transformers](https://huggingface.co/docs/transformers/en/index) documentation.\n",
     "\n",
-    "> **Note:** Gemma 3 and later models support input of both images and text. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../get-started/index.md).\n",
+    "> **Note:** Gemma 3 and later models support input of both images and text. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../models/paligemma/index.md).\n",
     "\n",
     "\n",
     "### Install Python packages\n",
diff --git a/docs/notebooks/run-gemma-with-keras-gemma-4-26b-a4b.ipynb b/docs/notebooks/run-gemma-with-keras-gemma-4-26b-a4b.ipynb
index e229a9c..1947409 100644
--- a/docs/notebooks/run-gemma-with-keras-gemma-4-26b-a4b.ipynb
+++ b/docs/notebooks/run-gemma-with-keras-gemma-4-26b-a4b.ipynb
@@ -37,7 +37,7 @@
     "\n",
     "Generating content, summarizing, and analysing content are just some of the tasks you can accomplish with Gemma open models. This tutorial shows you how to get started running Gemma using Keras, including generating text content with text and image input. [Keras](https://keras.io/) provides implementations for running Gemma and other models using JAX, PyTorch, and TensorFlow. If you're new to Keras, you might want to read [Getting started with Keras](https://keras.io/getting_started/) before you begin.\n",
     "\n",
-    "Gemma 3 and later models support text and image input. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../get-started/index.md).\n",
+    "Gemma 3 and later models support text and image input. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../models/paligemma/index.md).\n",
     "> **Note:** This notebook won't run on T4 if you try with Gemma 4. Use L4 or higher GPUs.\n",
     "\n",
     "### Install Keras packages\n",
diff --git a/docs/notebooks/run-gemma-with-keras-gemma-4-31b.ipynb b/docs/notebooks/run-gemma-with-keras-gemma-4-31b.ipynb
index 6c492b3..59b592d 100644
--- a/docs/notebooks/run-gemma-with-keras-gemma-4-31b.ipynb
+++ b/docs/notebooks/run-gemma-with-keras-gemma-4-31b.ipynb
@@ -37,7 +37,7 @@
     "\n",
     "Generating content, summarizing, and analysing content are just some of the tasks you can accomplish with Gemma open models. This tutorial shows you how to get started running Gemma using Keras, including generating text content with text and image input. [Keras](https://keras.io/) provides implementations for running Gemma and other models using JAX, PyTorch, and TensorFlow. If you're new to Keras, you might want to read [Getting started with Keras](https://keras.io/getting_started/) before you begin.\n",
     "\n",
-    "Gemma 3 and later models support text and image input. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../get-started/index.md).\n",
+    "Gemma 3 and later models support text and image input. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../models/paligemma/index.md).\n",
     "> **Note:** This notebook won't run on T4 if you try with Gemma 4. Use L4 or higher GPUs.\n",
     "\n",
     "### Install Keras packages\n",
diff --git a/docs/notebooks/run-gemma-with-keras-gemma-4-e2b.ipynb b/docs/notebooks/run-gemma-with-keras-gemma-4-e2b.ipynb
index 8840fd8..448098c 100644
--- a/docs/notebooks/run-gemma-with-keras-gemma-4-e2b.ipynb
+++ b/docs/notebooks/run-gemma-with-keras-gemma-4-e2b.ipynb
@@ -37,7 +37,7 @@
     "\n",
     "Generating content, summarizing, and analysing content are just some of the tasks you can accomplish with Gemma open models. This tutorial shows you how to get started running Gemma using Keras, including generating text content with text and image input. [Keras](https://keras.io/) provides implementations for running Gemma and other models using JAX, PyTorch, and TensorFlow. If you're new to Keras, you might want to read [Getting started with Keras](https://keras.io/getting_started/) before you begin.\n",
     "\n",
-    "Gemma 3 and later models support text and image input. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../get-started/index.md).\n",
+    "Gemma 3 and later models support text and image input. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../models/paligemma/index.md).\n",
     "> **Note:** This notebook won't run on T4 if you try with Gemma 4. Use L4 or higher GPUs.\n",
     "\n",
     "### Install Keras packages\n",
diff --git a/docs/notebooks/run-gemma-with-keras-gemma-4-e4b.ipynb b/docs/notebooks/run-gemma-with-keras-gemma-4-e4b.ipynb
index 0a67c93..2cc75bf 100644
--- a/docs/notebooks/run-gemma-with-keras-gemma-4-e4b.ipynb
+++ b/docs/notebooks/run-gemma-with-keras-gemma-4-e4b.ipynb
@@ -37,7 +37,7 @@
     "\n",
     "Generating content, summarizing, and analysing content are just some of the tasks you can accomplish with Gemma open models. This tutorial shows you how to get started running Gemma using Keras, including generating text content with text and image input. [Keras](https://keras.io/) provides implementations for running Gemma and other models using JAX, PyTorch, and TensorFlow. If you're new to Keras, you might want to read [Getting started with Keras](https://keras.io/getting_started/) before you begin.\n",
     "\n",
-    "Gemma 3 and later models support text and image input. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../get-started/index.md).\n",
+    "Gemma 3 and later models support text and image input. Earlier versions of Gemma only support text input, except for some variants, including [PaliGemma](../models/paligemma/index.md).\n",
     "> **Note:** This notebook won't run on T4 if you try with Gemma 4. Use L4 or higher GPUs.\n",
     "\n",
     "### Install Keras packages\n",
diff --git a/docs/resources/embedding-draw-challenge.md b/docs/resources/embedding-draw-challenge.md
index c728eb8..a6fd4af 100644
--- a/docs/resources/embedding-draw-challenge.md
+++ b/docs/resources/embedding-draw-challenge.md
@@ -1,7 +1,7 @@
 ---
 title: Embedding Draw Challenge
 description: Sketch target objects on an interactive canvas and see how EmbeddingGemma 2 scores cross-modal image-to-text similarity in real time.
-models: [embeddinggemma]
+models: [embeddinggemma-2]
 hide:
   - toc
 ---
diff --git a/docs/stylesheets/motion.css b/docs/stylesheets/motion.css
index 2526c0b..63c2074 100644
--- a/docs/stylesheets/motion.css
+++ b/docs/stylesheets/motion.css
@@ -44,11 +44,11 @@
 
 .md-content__inner,
 .gemma-stage {
-  animation: gemmaPageReveal 260ms var(--gemma-ease) both;
+  animation: gemmaPageReveal 260ms var(--gemma-ease) backwards;
 }
 
 .md-sidebar--primary .md-nav--primary > .md-nav__list {
-  animation: gemmaPageReveal 220ms var(--gemma-ease) both;
+  animation: gemmaPageReveal 220ms var(--gemma-ease) backwards;
 }
 
 /* -------------------------------------------------------------------
diff --git a/docs/tuning/audio-qlora.md b/docs/tuning/audio-qlora.md
index 6f02944..2c7d2d8 100644
--- a/docs/tuning/audio-qlora.md
+++ b/docs/tuning/audio-qlora.md
@@ -258,8 +258,8 @@ trainable params: 48,316,416 || all params: 5,152,613,920 || trainable%: 0.9377
 
 To ensure the model learns strictly to transcribe the audio rather than memorizing prompt tokens:
 
-- The following code parses <code><span class="token-turn"><|turn></span><span class="token-role">model</span>\n</code> in each sequence.
-- All tokens prior to <code><span class="token-turn"><|turn></span><span class="token-role">model</span>\n</code>  (including audio prefix markers and user instruction) are masked to `-100`.
+- The following code parses <code><span class="token-turn">&lt;|turn&gt;</span><span class="token-role">model</span>\n</code> in each sequence.
+- All tokens prior to <code><span class="token-turn">&lt;|turn&gt;</span><span class="token-role">model</span>\n</code>  (including audio prefix markers and user instruction) are masked to `-100`.
 - Only assistant response tokens contribute to the cross-entropy loss.
 
 
diff --git a/docs/tutorials/gemma-explained/paligemma.md b/docs/tutorials/gemma-explained/paligemma.md
index cc9dc6e..dd67e8d 100644
--- a/docs/tutorials/gemma-explained/paligemma.md
+++ b/docs/tutorials/gemma-explained/paligemma.md
@@ -135,7 +135,7 @@ It takes as input the multi-modal representation from the projector and generate
 
 For the text input, each checkpoint was trained with various sequence lengths. For example, paligemma-3b-mix-224 was trained with sequence length 256 (input text + output text tokenized by Gemma’s tokenizer).
 
-PaliGemma uses the Gemma tokenizer with 256000 tokens, but extends its vocabulary with 1024 entries that represent coordinates in normalized image-space (<loc0000>...<loc1023>), and another with 128 entries (<seg000>...<seg127>) that are codewords used by a lightweight referring-expression segmentation [vector-quantized variational auto-encoder (VQ-VAE)](https://arxiv.org/abs/2301.02229). (256000 + 1024 + 128 = 257216)
+PaliGemma uses the Gemma tokenizer with 256000 tokens, but extends its vocabulary with 1024 entries that represent coordinates in normalized image-space (`<loc0000>`...`<loc1023>`), and another with 128 entries (`<seg000>`...`<seg127>`) that are codewords used by a lightweight referring-expression segmentation [vector-quantized variational auto-encoder (VQ-VAE)](https://arxiv.org/abs/2301.02229). (256000 + 1024 + 128 = 257216)
 
 ## Object Segmentation Example
 
diff --git a/docs/tutorials/visual-guides/diffusiongemma.md b/docs/tutorials/visual-guides/diffusiongemma.md
index b953f65..76f4396 100644
--- a/docs/tutorials/visual-guides/diffusiongemma.md
+++ b/docs/tutorials/visual-guides/diffusiongemma.md
@@ -11,7 +11,7 @@ source_url: https://newsletter.maartengrootendorst.com/p/a-visual-guide-to-diffu
     Originally published on
     [Exploring Language Models](https://newsletter.maartengrootendorst.com/p/a-visual-guide-to-diffusiongemma){ data-model="diffusiongemma" }.
 
-Following the release of [Gemma 4 12B](https://newsletter.maartengrootendorst.com/p/a-visual-guide-to-gemma-4-12b){ data-model="gemma-4" } there is yet another innovative model that I can introduce to you all, [DiffusionGemma](https://huggingface.co/google/diffusiongemma-26B-A4B-it){ data-model="diffusiongemma" }!
+Following the release of [Gemma 4 12B](gemma-4-12b.md){ data-model="gemma-4" } there is yet another innovative model that I can introduce to you all, [DiffusionGemma](https://huggingface.co/google/diffusiongemma-26B-A4B-it){ data-model="diffusiongemma" }!
 
 This is one I’m especially excited about because it’s a different way of approaching text generation, yet uses many of the great things we have with existing “regular” LLMs. We are going to explore the old, the new, and how they live side-by-side.
 
@@ -198,7 +198,7 @@ So how can we have the benefits of Uniform State Diffusion without the downsides
 ## Architecture of DiffusionGemma
 **DiffusionGemma** applies Uniform State Diffusion that we explore thus far. However, the downsides of training this entirely from scratch still remains. The solution was simply not to train from scratch but to use an existing checkpoint as a start instead, namely the Gemma 4 26B A4B model.
 
-Gemma 4 26B A4B, as covered in [A Visual Guide to Gemma 4](https://newsletter.maartengrootendorst.com/p/a-visual-guide-to-gemma-4){ data-model="gemma-4" }, is a Mixture of Experts (MoE) model that already went through extensive training and has great performance. There is one problem though…
+Gemma 4 26B A4B, as covered in [A Visual Guide to Gemma 4](gemma-4.md){ data-model="gemma-4" }, is a Mixture of Experts (MoE) model that already went through extensive training and has great performance. There is one problem though…
 
 Gemma 4 26B A4B is a *decoder-only* model and meant to generate text one by one whilst diffusion generally needs an *encoder* and a *denoiser*. This is where an elegant solution comes in, the **Encoder-Denoiser** patch!
 
diff --git a/docs/tutorials/visual-guides/gemma-4-12b.md b/docs/tutorials/visual-guides/gemma-4-12b.md
index 326db90..58eff78 100644
--- a/docs/tutorials/visual-guides/gemma-4-12b.md
+++ b/docs/tutorials/visual-guides/gemma-4-12b.md
@@ -150,17 +150,17 @@ Gemma 4 12B is not just a model that fits nicely between E4B and 26B A4B, it doe
 ## What's next
 
 <div class="card-grid">
-<a class="card" href="mtp">
+<a class="card" href="mtp.md">
   <span class="card-icon">{{ icon('lucide/book-open') }}</span>
   <div class="card-title">A Visual Guide to MTP</div>
   <p class="card-desc">Drafter models and speculative decoding, illustrated.</p>
 </a>
-<a class="card" href="diffusiongemma">
+<a class="card" href="diffusiongemma.md">
   <span class="card-icon">{{ icon('lucide/book-open') }}</span>
   <div class="card-title">A Visual Guide to DiffusionGemma</div>
   <p class="card-desc">How denoising produces text instead of pixels.</p>
 </a>
-<a class="card" href="../../resources">
+<a class="card" href="../../resources/index.md">
   <span class="card-icon">{{ icon('lucide/library') }}</span>
   <div class="card-title">Resources</div>
   <p class="card-desc">Demos, tools and the repositories worth bookmarking.</p>
diff --git a/scripts/agent_docs.py b/scripts/agent_docs.py
index 81916f5..95ac43a 100644
--- a/scripts/agent_docs.py
+++ b/scripts/agent_docs.py
@@ -26,6 +26,7 @@
 import re
 import sys
 import tomllib
+from functools import lru_cache
 from pathlib import Path
 from typing import Iterable
 
@@ -37,6 +38,7 @@
 CONFIG_FILE = REPO_ROOT / "zensical.toml"
 MODELS_FILE = REPO_ROOT / "data" / "models.yml"
 FRAMEWORKS_FILE = REPO_ROOT / "data" / "frameworks.yml"
+LINKS_FILE = REPO_ROOT / "data" / "links.yml"
 
 # Pages call `icon()`; the macros extension provides it during a Zensical build, but these raw
 # routes are rendered here, in a Jinja environment of our own. Same source, same output.
@@ -56,6 +58,7 @@
 RELATIVE_ASSET = re.compile(
     r"(?P<prefix>\]\(|(?:\b(?:src|href)=[\"']))(?:\.\.?/|/)*(?P<path>assets/[^)\s\"']+)"
 )
+REFERENCE_LINK = re.compile(r"\[([^\]]+)\]\[([a-z0-9][a-z0-9-]*)\]")
 
 
 def split_frontmatter(text: str) -> tuple[dict, str]:
@@ -121,13 +124,24 @@ def markdown_route_for(relative: str) -> str:
     return "/index.md" if route == "/" else f"{route.rstrip('/')}.md"
 
 
-def display_title(metadata: dict, family_names: dict[str, str]) -> str:
+def display_title(
+    metadata: dict,
+    family_names: dict[str, str],
+    relative: str = "",
+) -> str:
     title = str(metadata.get("title", "Untitled"))
+    if not relative.startswith("models/"):
+        return title
     models = metadata.get("models") or []
-    if title.lower() in {"model card", "overview"} and len(models) == 1:
-        family = str(models[0]).split("/", 1)[0]
-        if family in family_names:
-            return f"{family_names[family]} — {title}"
+    parts = Path(relative).parts
+    family = (
+        parts[1]
+        if len(parts) > 2 and parts[1] in family_names
+        else (str(models[0]).split("/", 1)[0] if len(models) == 1 else "")
+    )
+    family_name = family_names.get(family)
+    if family_name and not all(part in title.lower() for part in family_name.lower().split()):
+        return f"{family_name} — {title}"
     return title
 
 
@@ -139,6 +153,12 @@ def framework_data() -> dict:
     return yaml.safe_load(FRAMEWORKS_FILE.read_text(encoding="utf-8"))
 
 
+@lru_cache(maxsize=1)
+def shared_links() -> dict[str, str]:
+    data = yaml.safe_load(LINKS_FILE.read_text(encoding="utf-8"))
+    return {name: str(entry["url"]) for name, entry in (data.get("links") or {}).items()}
+
+
 def render_llms() -> str:
     pages = source_pages()
     data = model_data()
@@ -159,7 +179,7 @@ def render_llms() -> str:
         for relative in entries:
             seen.add(relative)
             _path, metadata, _body = pages[relative]
-            title = display_title(metadata, family_names)
+            title = display_title(metadata, family_names, relative)
             description = " ".join(str(metadata.get("description", "")).split())
             url = f"{SITE_URL}{markdown_route_for(relative)}"
             suffix = f": {description}" if description else ""
@@ -185,6 +205,12 @@ def render_raw(relative: str, metadata: dict, body: str) -> str:
     context["frameworkdata"] = framework_data()
     context["icon"] = _icon
     rendered = environment.from_string(body).render(**context)
+
+    def resolve_reference(match: re.Match) -> str:
+        url = shared_links().get(match.group(2))
+        return f"[{match.group(1)}]({url})" if url else match.group(0)
+
+    rendered = REFERENCE_LINK.sub(resolve_reference, rendered)
     rendered = WHATS_NEXT_SECTION.sub("", rendered)
     rendered = SNIPPET_INCLUDE.sub("", rendered).strip() + "\n"
     rendered = RELATIVE_ASSET.sub(rf"\g<prefix>{SITE_URL}/\g<path>", rendered)
diff --git a/scripts/frontmatter_schema.py b/scripts/frontmatter_schema.py
index b24c828..8153cf3 100644
--- a/scripts/frontmatter_schema.py
+++ b/scripts/frontmatter_schema.py
@@ -94,8 +94,9 @@ class Field:
 
 @cache
 def model_families() -> frozenset[str]:
-    """Every family id. Derived from `docs/models/`, so adding a model directory is enough."""
-    return frozenset(p.name for p in (DOCS_DIR / "models").iterdir() if p.is_dir())
+    """Every family id in data/models.yml."""
+    data = yaml.safe_load(MODELS_DATA.read_text(encoding="utf-8"))
+    return frozenset(str(family["id"]) for family in data.get("families", []))
 
 
 @cache
diff --git a/scripts/lint_docs.py b/scripts/lint_docs.py
index 01e81da..4a4a410 100644
--- a/scripts/lint_docs.py
+++ b/scripts/lint_docs.py
@@ -231,23 +231,26 @@ def load_context() -> Context:
     return Context(pages=load_pages(), nav_entries=list(walk_nav(nav)), config=config)
 
 
+FENCE_OPEN = re.compile(r"^(\s*)(`{3,}|~{3,})(.*)$")
+
+
 def iter_code_blocks(page: Page) -> Iterator[tuple[int, str, str]]:
     """Yield (start line, info string, block content) for each fenced code block."""
-    fence = re.compile(r"^(\s*)```(.*)$")
     open_at: int | None = None
+    active_fence = ""
     info = ""
     buf: list[str] = []
     for lineno, line in page.body_lines():
-        m = fence.match(line)
-        if not m:
-            if open_at is not None:
-                buf.append(line)
-            continue
+        m = FENCE_OPEN.match(line)
         if open_at is None:
-            open_at, info, buf = lineno, m.group(2).strip(), []
-        else:
+            if m:
+                open_at, active_fence, info, buf = lineno, m.group(2), m.group(3).strip(), []
+            continue
+        if m and m.group(2) == active_fence and not m.group(3).strip():
             yield open_at, info, "\n".join(buf)
             open_at = None
+        else:
+            buf.append(line)
 
 
 INLINE_CODE = re.compile(r"`+[^`]*`+")
@@ -264,13 +267,17 @@ def strip_inline_code(line: str) -> str:
 
 def outside_code(page: Page) -> Iterator[tuple[int, str]]:
     """Yield body lines that are not inside a fenced code block, with code spans blanked."""
-    in_fence = False
+    active_fence: str | None = None
     for lineno, line in page.body_lines():
-        if line.lstrip().startswith("```"):
-            in_fence = not in_fence
-            continue
-        if not in_fence:
+        m = FENCE_OPEN.match(line)
+        if active_fence is None:
+            if m:
+                active_fence = m.group(2)
+                continue
             yield lineno, strip_inline_code(line)
+        elif m and m.group(2) == active_fence and not m.group(3).strip():
+            active_fence = None
+            continue
 
 
 # --- Checks -----------------------------------------------------------------------------------
@@ -558,9 +565,19 @@ def check_model_swap(ctx: Context) -> Iterable[Finding]:
     """A swappable engine must not also appear in unacknowledged static code blocks."""
     data = yaml.safe_load((REPO_ROOT / "data" / "models.yml").read_text(encoding="utf-8"))
     engines: dict[str, set[str]] = {}
+    index: list[tuple[str, str, str]] = []
     for size in data.get("gemma4", {}).get("sizes", []):
+        size_id = str(size["id"])
         for engine, identifier in size.get("ids", {}).items():
             engines.setdefault(str(engine), set()).add(str(identifier))
+            index.append((str(identifier), str(engine), size_id))
+    index.sort(key=lambda item: len(item[0]), reverse=True)
+    pattern = re.compile("|".join(re.escape(text) for text, _, _ in index))
+    by_text = {text: (engine, size_id) for text, engine, size_id in index}
+    unrewritten_re = re.compile(
+        r"\bgemma[-_]?4[-_:(](?:e2b|e4b|12b|26b(?:[-_]a4b)?|31b)[\w.-]*",
+        re.IGNORECASE,
+    )
 
     for page in ctx.pages:
         blocks = list(iter_code_blocks(page))
@@ -568,12 +585,40 @@ def check_model_swap(ctx: Context) -> Iterable[Finding]:
         if not marked:
             continue
 
-        active = {
-            engine
-            for _, _, code in marked
-            for engine, identifiers in engines.items()
-            if any(identifier in code for identifier in identifiers)
-        }
+        active: set[str] = set()
+        page_default_size: str | None = None
+        for line, _, code in marked:
+            matches = [m.group(0) for m in pattern.finditer(code)]
+            if not matches:
+                yield Finding(
+                    "model-swap",
+                    page.repo_rel,
+                    line,
+                    ".model-swap block contains no swappable model identifier from data/models.yml",
+                )
+                continue
+            for match in matches:
+                engine, size_id = by_text[match]
+                active.add(engine)
+                if page_default_size is None:
+                    page_default_size = size_id
+                elif size_id != page_default_size:
+                    yield Finding(
+                        "model-swap",
+                        page.repo_rel,
+                        line,
+                        f".model-swap identifier '{match}' targets '{size_id}' "
+                        f"but page default is '{page_default_size}'",
+                    )
+            remainder = pattern.sub("", code)
+            for leftover in unrewritten_re.finditer(remainder):
+                yield Finding(
+                    "model-swap",
+                    page.repo_rel,
+                    line,
+                    f"unrewritten model identifier '{leftover.group(0)}' inside .model-swap block",
+                )
+
         for line, info, code in blocks:
             if ".model-swap" in info or ".model-fixed" in info or ".output" in info:
                 continue
diff --git a/scripts/notebooks.py b/scripts/notebooks.py
index bc88aa9..064fe19 100644
--- a/scripts/notebooks.py
+++ b/scripts/notebooks.py
@@ -73,6 +73,16 @@
     r"<!--\s*notebook:skip-start\s*-->.*?<!--\s*notebook:skip-end\s*-->",
     re.DOTALL,
 )
+REQUIRES_REGION = re.compile(
+    r"<!--\s*notebook:requires-start\s+(?P<modality>[a-z0-9_-]+)\s*-->[ \t]*\n?"
+    r"(?P<content>.*?)"
+    r"<!--\s*notebook:requires-end\s*-->[ \t]*\n?",
+    re.DOTALL,
+)
+REQUIRES_COMMENT = re.compile(
+    r"<!--\s*notebook:requires-(?:start\s+[a-z0-9_-]+|end)\s*-->[ \t]*\n?"
+)
+FENCE_REQUIRES = re.compile(r"\b(?:data-)?requires=[\"'](?P<modality>[a-z0-9_-]+)[\"']")
 ADMONITION = re.compile(
     r"^!!!\s+(?P<label>[A-Za-z]+)(?:\s+\"[^\"]*\")?\n(?P<body>(?:[ \t]+.*(?:\n|$))*)",
     re.MULTILINE,
@@ -196,13 +206,16 @@ def extract_swappable_blocks(body: str) -> list[str]:
 
 def detect_engines_in_use(body: str) -> set[str]:
     index = get_model_index()
+    if not index:
+        return set()
+    text_to_engine = {text: engine for text, engine, _ in index}
+    pattern = re.compile("|".join(re.escape(text) for text, _, _ in index))
     blocks = extract_swappable_blocks(body)
-    engines = set()
-    for block in blocks:
-        for text, engine, _ in index:
-            if text in block:
-                engines.add(engine)
-    return engines
+    return {
+        text_to_engine[match.group(0)]
+        for block in blocks
+        for match in pattern.finditer(block)
+    }
 
 
 def candidate_models(source: Path) -> list[str]:
@@ -238,8 +251,28 @@ def repl(match: re.Match) -> str:
     return pattern.sub(repl, block)
 
 
+def model_supports_modality(model: str | None, modality: str) -> bool:
+    if not model:
+        return True
+    models = get_gemma4_models()
+    if model not in models:
+        return True
+    return modality in (models[model].get("inputs") or [])
+
+
+def filter_requires_regions(text: str, model: str | None = None) -> str:
+    def repl(match: re.Match) -> str:
+        if model_supports_modality(model, match.group("modality")):
+            return match.group("content")
+        return ""
+
+    text = REQUIRES_REGION.sub(repl, text)
+    return REQUIRES_COMMENT.sub("", text)
+
+
 def notebook_markdown(text: str) -> str:
     text = SKIP_REGION.sub("", text)
+    text = REQUIRES_COMMENT.sub("", text)
 
     def resolve_reference(match: re.Match) -> str:
         url = shared_links().get(match.group(2))
@@ -294,6 +327,7 @@ def code_cell(language: str, source: str) -> dict:
 
 def export_cells(body: str, model: str | None = None) -> list[dict]:
     body = SKIP_REGION.sub("", body)
+    body = filter_requires_regions(body, model=model)
     lines = body.splitlines(keepends=True)
     cells: list[dict] = []
     prose: list[str] = []
@@ -312,6 +346,10 @@ def export_cells(body: str, model: str | None = None) -> list[dict]:
             raise ValueError(f"unclosed code fence at line {index + 1}")
 
         attributes = match.group("attributes")
+        req_match = FENCE_REQUIRES.search(attributes)
+        if req_match and not model_supports_modality(model, req_match.group("modality")):
+            index = closing + 1
+            continue
         block = "".join(lines[index + 1 : closing])
         if model and ".model-swap" in attributes:
             block = rewrite_block(block, model)
@@ -387,7 +425,7 @@ def render(
     if target_name is None:
         stem = default_stem(source)
         if model:
-            target_name = f"{stem}-{model.replace('/', '-')}-it.ipynb"
+            target_name = f"{stem}-{model.replace('/', '-')}.ipynb"
         else:
             target_name = f"{stem}.ipynb"
 
diff --git a/snippets/links.md b/snippets/links.md
index 9227385..8b5ed33 100644
--- a/snippets/links.md
+++ b/snippets/links.md
@@ -2,7 +2,7 @@
 
 [responsible-ai]: https://ai.google.dev/responsible
 [gemma-terms]: /resources/terms
-[prohibited-use]: /resources/prohibited_use_policy
+[prohibited-use]: /resources/prohibited-use
 [hf-gemma4]: https://huggingface.co/collections/google/gemma-4
 [gemma4-hf-collection]: https://huggingface.co/collections/google/gemma-4
 [kaggle-gemma4]: https://www.kaggle.com/models/google/gemma-4
@@ -35,7 +35,7 @@
 [recurrentgemma-release-hf-collection]: https://huggingface.co/collections/google/recurrentgemma-release
 [codegemma-release-hf-collection]: https://huggingface.co/collections/google/codegemma-release
 [gemmaverse]: https://deepmind.google/models/gemma/gemmaverse/
-[gemma-cookbook]: https://github.com/google-gemma/gemma-cookbook
+[gemma-cookbook]: https://github.com/google-gemma/cookbook
 [gemma-skills-repo]: https://github.com/google-gemma/gemma-skills
 [awesome-gemma]: https://github.com/google-gemma/awesome-gemma
 [gemma-translator]: https://github.com/google-gemma/gemma-translator
diff --git a/tests/test_agent_docs.py b/tests/test_agent_docs.py
index 0fec03c..10a0d4f 100644
--- a/tests/test_agent_docs.py
+++ b/tests/test_agent_docs.py
@@ -41,9 +41,70 @@ def test_llms_uses_absolute_markdown_links(self):
         rendered = agent_docs.render_llms()
         self.assertTrue(rendered.startswith("# Gemma Documentation\n"))
         self.assertIn("https://gemma.ai/get-started.md", rendered)
+        self.assertIn("- [Overview](https://gemma.ai/deployment.md):", rendered)
+        self.assertNotIn("[Gemma 4 — Overview](https://gemma.ai/deployment.md)", rendered)
+        self.assertIn(
+            "- [DiffusionGemma — Generate output](https://gemma.ai/models/diffusiongemma/inference.md):",
+            rendered,
+        )
+        self.assertIn(
+            "- [RecurrentGemma — Fine-tuning using JAX and Flax](https://gemma.ai/models/recurrentgemma/fine-tuning.md):",
+            rendered,
+        )
         self.assertNotIn("llms-full", rendered)
         self.assertNotIn("/contributing/", rendered)
 
+    def test_display_title_scopes_family_prefix_to_model_pages(self):
+        family_names = {
+            "gemma-4": "Gemma 4",
+            "diffusiongemma": "DiffusionGemma",
+            "recurrentgemma": "RecurrentGemma",
+            "gemma-aps": "Gemma APS",
+        }
+        self.assertEqual(
+            agent_docs.display_title(
+                {"title": "Overview", "models": ["gemma-4"]},
+                family_names,
+                "deployment/index.md",
+            ),
+            "Overview",
+        )
+        self.assertEqual(
+            agent_docs.display_title(
+                {"title": "Model card", "models": ["gemma-4"]},
+                family_names,
+                "models/gemma-4/model-card.md",
+            ),
+            "Gemma 4 — Model card",
+        )
+        self.assertEqual(
+            agent_docs.display_title(
+                {"title": "Generate output", "models": ["diffusiongemma"]},
+                family_names,
+                "models/diffusiongemma/inference.md",
+            ),
+            "DiffusionGemma — Generate output",
+        )
+        self.assertEqual(
+            agent_docs.display_title(
+                {"title": "Fine-tuning using JAX and Flax", "models": ["recurrentgemma"]},
+                family_names,
+                "models/recurrentgemma/fine-tuning.md",
+            ),
+            "RecurrentGemma — Fine-tuning using JAX and Flax",
+        )
+        self.assertEqual(
+            agent_docs.display_title(
+                {
+                    "title": "Gemma for abstractive proposition segmentation (APS)",
+                    "models": ["gemma-aps"],
+                },
+                family_names,
+                "models/gemma-aps/index.md",
+            ),
+            "Gemma for abstractive proposition segmentation (APS)",
+        )
+
     def test_raw_markdown_renders_model_macros(self):
         pages = agent_docs.source_pages()
         _path, metadata, body = pages["get-started/index.md"]
@@ -65,6 +126,26 @@ def test_raw_markdown_resolves_relative_asset_paths(self):
             rendered,
         )
 
+    def test_raw_markdown_resolves_shared_reference_links(self):
+        pages = agent_docs.source_pages()
+        _path, metadata, body = pages["models/gemma-4/index.md"]
+        rendered = agent_docs.render_raw("models/gemma-4/index.md", metadata, body)
+        self.assertIn(
+            "[Hugging Face](https://huggingface.co/collections/google/gemma-4)",
+            rendered,
+        )
+        self.assertIn("[commercial use](../../resources/apache-2.md)", rendered)
+        self.assertNotIn("[hf-gemma4]", rendered)
+        self.assertNotIn("[gemma-terms]", rendered)
+
+        _path, metadata, body = pages["models/gemma-3/index.md"]
+        rendered = agent_docs.render_raw("models/gemma-3/index.md", metadata, body)
+        self.assertIn(
+            "[Gemma Prohibited Use Policy](/resources/prohibited-use)",
+            rendered,
+        )
+        self.assertNotIn("[prohibited-use]", rendered)
+
     def test_raw_targets_include_inline_text_routes(self):
         site_dir = Path("/tmp/site")
         targets = {path.relative_to(site_dir).as_posix() for path, _ in agent_docs.raw_targets(site_dir)}
@@ -73,6 +154,20 @@ def test_raw_targets_include_inline_text_routes(self):
         self.assertIn("get-started/best-practices.md.txt", targets)
         self.assertFalse(any(target.startswith("contributing/") for target in targets))
 
+    def test_variant_generation_ids_accepted_in_frontmatter(self):
+        import frontmatter_schema
+
+        for family_id in ("embeddinggemma-2", "paligemma-2", "shieldgemma-2"):
+            self.assertIn(family_id, frontmatter_schema.model_families())
+            problems = frontmatter_schema.validate(
+                {
+                    "title": "Test Page",
+                    "description": "A valid description string that is between forty and two hundred characters.",
+                    "models": [family_id],
+                }
+            )
+            self.assertEqual(problems, [])
+
 
 if __name__ == "__main__":
     unittest.main()
diff --git a/tests/test_notebooks.py b/tests/test_notebooks.py
index 5ae6426..cdfb99c 100644
--- a/tests/test_notebooks.py
+++ b/tests/test_notebooks.py
@@ -130,6 +130,20 @@ def test_audio_tutorial_filters_models(self):
         self.assertNotIn("gemma-4/26b-a4b", candidates)
         self.assertNotIn("gemma-4/31b", candidates)
 
+    def test_requires_region_and_cell_filtered_by_model_modality(self):
+        hf_source = notebooks.REPO_ROOT / "docs" / "deployment" / "huggingface.md"
+        for audio_model in ("gemma-4/e2b", "gemma-4/e4b", "gemma-4/12b"):
+            payload = json.loads(notebooks.render(hf_source, model=audio_model))
+            code_sources = "\n".join(
+                "".join(cell["source"]) for cell in payload["cells"][2:] if cell["cell_type"] == "code"
+            )
+            self.assertIn("roses-are.wav", code_sources)
+        for non_audio_model in ("gemma-4/26b-a4b", "gemma-4/31b"):
+            payload = json.loads(notebooks.render(hf_source, model=non_audio_model))
+            all_sources = "\n".join("".join(cell["source"]) for cell in payload["cells"][2:])
+            self.assertNotIn("roses-are.wav", all_sources)
+            self.assertNotIn("Generate text from audio data", all_sources)
+
     def test_infer_model_from_output_filename(self):
         self.assertEqual(notebooks.infer_model("thinking-gemma-4-e2b-it.ipynb"), "gemma-4/e2b")
         self.assertEqual(notebooks.infer_model("thinking-gemma-4-26b-a4b-it.ipynb"), "gemma-4/26b-a4b")