asher · Oct 18, 2026 · Oct 18, 2026 · Oct 18, 2026 · Oct 18, 2026 · Oct 18, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -8,6 +8,12 @@ adhere to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ### Added
 
+- `gmlx pull`, `image generate`, `serve` and the ComfyUI pack run MOSS-Image
+  and MOSS-Image Base GGUFs, and their LoRAs.
+- The ComfyUI pack runs MOSS-Image without the OpenMOSS plugin, and takes
+  `SkipLayerGuidanceDiTSimple` on it to leave blocks out of the negative pass.
+- MOSS-Image takes an `upsampler`, on a caption prompt gmlx writes from the
+  OpenMOSS prompt-enhancer skill.
 - FLUX.3 klein rewrites each prompt with a chat model of your choice before
   encoding it, set as the entry's `upsampler` or by `--upsampler`.
 - FLUX.2 klein, Krea 2, Z-Image and Qwen-Image 2.1 take an `upsampler` too,
@@ -53,6 +59,8 @@ adhere to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
   `--chat-template` as invalid Jinja.
 - `--thinking off` was silently ignored on a model with no thinking switch
   when its template has the word thinking in its text. It now warns.
+- A Krea 2 GGUF whose stored projector shape is not a list of integers
+  crashed the load. It is refused with the reason.
 - A streamed image request refused while it ran, such as one too large for
   memory, got a server error frame where the same request unstreamed got a 400.
 - The `gmlx pull` plan of an image model could say its parts fit in stages

diff --git a/README.md b/README.md
@@ -98,8 +98,8 @@ The recording runs at true speed, with a 27B model in a local server.
 - Embeddings, rerank, speech-to-text and text-to-speech on the same server,
   for a local RAG and voice stack. See
   [Speech, embeddings and rerank](https://asher.github.io/gmlx/services.html).
-- Image generation from Z-Image, Qwen-Image 2.1, Krea 2 Turbo, FLUX.2 klein
-  and FLUX.3 klein GGUFs, on the command line, on the OpenAI image routes
+- Image generation from Z-Image, Qwen-Image 2.1, Krea 2 Turbo, MOSS-Image,
+  FLUX.2 klein and FLUX.3 klein GGUFs, on the command line, on the OpenAI image routes
   and in ComfyUI. It needs an M3 chip or newer. See
   [Image generation](https://asher.github.io/gmlx/images.html).
 - A decision API at `/v1/systemone` that returns a probability for each

diff --git a/THIRD_PARTY_NOTICES.md b/THIRD_PARTY_NOTICES.md
@@ -25,6 +25,9 @@ in [`licenses/`](licenses/).
 | [flux2](https://github.com/black-forest-labs/flux2) | Apache-2.0 | BFL's FLUX.2 reference at commit `50fe516`, ported to MLX and modified for gmlx: the DiT constants of `Klein4BParams`, `Klein9BParams` and `Flux2Params` and the tensor names and shapes of `Flux2` in `src/flux2/model.py`, the sampling defaults of `FLUX2_MODEL_INFO` in `src/flux2/util.py`, the prompt rules of `Qwen3Embedder` in `src/flux2/text_encoder.py`, the rounding of `denoise` and `denoise_cfg` and the reference image rules of `encode_image_refs` in `src/flux2/sampling.py` in `gmlx/image/families/flux2_klein/family.py`, the constants of `Klein9BParams`, the 9B rows of `FLUX2_MODEL_INFO` and the Qwen3-8B encoder of `load_qwen3_embedder` in `src/flux2/text_encoder.py` in `gmlx/image/families/flux2_klein_9b/family.py`, the `flux.2-klein-9b-kv` row of `FLUX2_MODEL_INFO` in `gmlx/image/families/flux2_klein_9b_kv/family.py` and the step of `denoise_cached` in `src/flux2/sampling.py` in `gmlx/image/families/flux2_klein_9b_kv/step.py`, the DiT forward of `src/flux2/model.py`, with `forward_kv_extract`, `forward_kv_cached` and `causal_attn_fn`, and the token ids of `prc_img`, `prc_txt` and `encode_image_refs` in `src/flux2/sampling.py` in `gmlx/image/families/flux2_klein/model.py`, and the schedule of `compute_empirical_mu` and `get_schedule` in `src/flux2/sampling.py` in `flux2_mu` and `flux2_sigmas` of `gmlx/image/schedule.py`, and the reference image preparation of `encode_image_refs` and `default_prep` in `src/flux2/sampling.py` in `gmlx/image/families/flux2_klein/edit.py`; and the upsampler's text to image system message `SYSTEM_MESSAGE_UPSAMPLING_T2I` of `src/flux2/system_messages.py`, shipped unmodified as `gmlx/image/families/flux2_klein/system_message_upsampling_t2i.txt`, with the chat turn of `scripts/cli.py` in `gmlx/image/families/flux2_klein/family.py`. |
 | [Z-Image-Turbo Space](https://huggingface.co/spaces/Tongyi-MAI/Z-Image-Turbo) | apache-2.0, from the Space's card; `pe.py` carries no license header | At revision `768cb50`, the prompt enhancer's system prompt `prompt_template` of `pe.py`, shipped as its string in `gmlx/image/families/z_image/pe_prompt_template.txt`, with its use in `app.py` in `gmlx/image/families/z_image/family.py`. |
 | [krea-2](https://github.com/krea-ai/krea-2) | Apache-2.0 | At commit `db3984f`, the prompt expansion system prompt `docs/expansion.txt`, shipped unmodified in `gmlx/image/families/krea2/`, with its use of `docs/prompting.md` in `gmlx/image/families/krea2/family.py`. The Krea 2 weights are under their own license and do not ship with gmlx. |
+| [MOSS-Image](https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT) | Apache-2.0 | OpenMOSS-Team/MOSS-Image-SFT at revision `fc5de0d`, ported to MLX and modified for gmlx: the tensor names and shapes of `MOSSImageTransformer` in `pipeline_moss_image.py` and the constants of that file and of `transformer/config.json`, and the prompt template of `tokenizer/chat_template.jinja` with the token cap of the pipeline, in `gmlx/image/families/moss_image/family.py`, and the DiT forward of `pipeline_moss_image.py` in `gmlx/image/families/moss_image/model.py` (license text in [`licenses/MOSS-Image-LICENSE`](licenses/MOSS-Image-LICENSE)). |
+| [MOSS-Image ComfyUI plugin](https://github.com/OpenMOSS/MOSS-Image/tree/main/comfyui) | Apache-2.0 | The OpenMOSS ComfyUI plugin at commit `c1355e6`: the model settings of `comfyui/moss-image/loaders.py` (a flow model on the FLUX.2 latent format, sampling shift 1.0, timestep multiplier 1000, bfloat16 and float32 inference, memory usage factor 4.0), which the `MOSSImage` model class in `gmlx/comfyui_pack/moss_image.py` takes. |
+| [MOSS-Image prompt-enhancer skill](https://github.com/OpenMOSS/MOSS-Image/tree/main/skills/moss-image-prompt-enhancer) | Apache-2.0 | At commit `c1355e6`, the caption rules of `skills/moss-image-prompt-enhancer/SKILL.md`, restated and modified for gmlx as the system prompt `gmlx/image/families/moss_image/caption_prompt.txt`, with the skill's `{"caption": ...}` reply read in `gmlx/image/families/moss_image/family.py`. The skill's layout drafts and its tool are not used. |
 | [ComfyUI-GGUF](https://github.com/city96/ComfyUI-GGUF) | Apache-2.0 | The model detection keys of `convert.py` at commit `6ea2651`, which `gmlx/image/fingerprint.py` uses to refuse other diffusion model layouts by name. |
 | [misaki](https://github.com/hexgrad/misaki) | Apache-2.0 | Whole package vendored at `gmlx/_vendor/misaki` (commit `fba1236`): Kokoro's grapheme-to-phoneme front-end, used by the TTS service until upstream's next PyPI release. |
 | [musl](https://musl.libc.org) | MIT | Statically linked into the guest entry binary `gmlx/container/guest/gmlx-entry` as its C library ([`licenses/musl-COPYRIGHT`](licenses/musl-COPYRIGHT)). |

diff --git a/docs/cli.md b/docs/cli.md
@@ -640,8 +640,8 @@ gmlx pull qwen-image --quant Q5_K_M
 
 An image model name, `z-image-turbo`, `qwen-image`, `qwen-image-turbo`,
 `krea-2-turbo`, `flux-2-klein`, `flux-2-klein-base`, `flux-2-klein-9b`,
-`flux-2-klein-base-9b`, `flux-2-klein-9b-kv`, `flux-3-klein` or
-`flux-3-klein-base`, pulls its DiT, text encoder and VAE, checks the VAE
+`flux-2-klein-base-9b`, `flux-2-klein-9b-kv`, `flux-3-klein`,
+`flux-3-klein-base`, `moss-image` or `moss-image-base`, pulls its DiT, text encoder and VAE, checks the VAE
 against its published sha256, packs it to GGUF, and adds an
 `image_models` entry. The quants default to
 the ones for this Mac's memory, and an encoder already on disk is used
@@ -971,6 +971,7 @@ the weights load.
 | Z-Image | Qwen3-4B | The FLUX.1 autoencoder | The variant's steps and guidance. |
 | Qwen-Image 2.1 | Qwen3-VL-8B, its text GGUF | The Qwen-Image 2.1 autoencoder | 40 steps and guidance 1. Turbo: 8 steps and guidance 1. |
 | Krea 2 | Qwen3-VL-4B-Instruct, its text GGUF | The Qwen-Image autoencoder | Turbo: 8 steps and guidance 1. |
+| MOSS-Image | Qwen3.5-9B | The MOSS-Image autoencoder | 50 steps and guidance 3, with block 22 left out of the negative pass. Base: the same. |
 | FLUX.3 klein | Qwen3-VL-4B-Instruct, its text GGUF | The FLUX.2 autoencoder | 16 steps and guidance 1. Base: 50 steps and guidance 3.5. |
 | FLUX.2 klein | Qwen3-4B | The FLUX.2 autoencoder | 4 steps and guidance 1, which is the only guidance it takes. Base: 50 steps and guidance 4. |
 | FLUX.2 klein 9B | Qwen3-8B | The FLUX.2 autoencoder | The same as FLUX.2 klein's. Its license allows non-commercial use only. |
@@ -1008,8 +1009,8 @@ loads from the encoder's mmproj GGUF, and the VAE encodes them. `--size`
 sets the size of the output, which defaults to the entry's `defaults.size`
 or the variant's, and not to a reference's. The fit check counts the
 references, and a request that does not fit is refused before any model
-loads. Z-Image and Krea 2 take no reference images, and `--ref` exits 2 on
-them.
+loads. Z-Image, Krea 2 and MOSS-Image take no reference images, and
+`--ref` exits 2 on them.
 
 `--ref` gives FLUX.3 klein up to 10 reference images too. Its text encoder
 reads the prompt alone, so it needs no `--encoder-mmproj`, and the VAE
@@ -1150,7 +1151,8 @@ and compares every tensor bit for bit before the file replaces `--out`.
 It packs the VAE of each image family: the FLUX.1 autoencoder for
 `z_image`, the Qwen-Image 2.1 autoencoder for `qwen_image21`, whose
 official file is F32 and stays F32, the Qwen-Image autoencoder for
-`krea2`, whose file is BF16, and the FLUX.2 autoencoder for `flux3_klein`,
+`krea2`, whose file is BF16, the MOSS-Image autoencoder for `moss_image`,
+whose file is F32, and the FLUX.2 autoencoder for `flux3_klein`,
 `flux2_klein` and `flux2_klein_9b`. BFL's FLUX.2 klein 4B file of that
 autoencoder is BF16 with its config, and the Comfy-Org
 `flux2-vae.safetensors` is F32 and needs no config. It packs from that

diff --git a/docs/comfyui-reference.md b/docs/comfyui-reference.md
@@ -74,8 +74,10 @@ ComfyUI sees each gmlx model as its own model class, so its stock nodes
 for that model work too, such as `ModelSamplingAuraFlow` and
 `TextEncodeQwenImage21`. The prompt nodes of the example workflows are
 ComfyUI's `CLIP Text Encode (Prompt)`, and the latent node is its
-`EmptySD3LatentImage`, or its `EmptyFlux2LatentImage` for FLUX.2 klein and
-FLUX.3 klein.
+`EmptySD3LatentImage`, or its `EmptyFlux2LatentImage` for FLUX.2 klein,
+FLUX.3 klein and MOSS-Image. ComfyUI has no model class for MOSS-Image,
+so the gmlx nodes register their own, with the settings of the OpenMOSS
+ComfyUI plugin, and stock nodes such as `ModelSamplingSD3` work on it.
 
 The sampler, the noise, the guidance and the schedule run in ComfyUI. Each
 call to the model is one request to the gmlx server, which runs the DiT,
@@ -105,6 +107,8 @@ that names it:
 - Hook keyframes and scheduled CLIP hooks.
 - A quantized Qwen-Image 2.1 prefix cache: `QwenImage21Cache` with a dtype
   other than the default.
+- A MOSS-Image conditioning that masks some of its text rows, such as one
+  from a text encoder other than `gmlx CLIP Loader`.
 - A FLUX.2 klein reference method other than `index`, such as `offset`
   from `FluxKontextMultiReferenceLatentMethod`, but for the
   `index_timestep_zero` method that `Flux KV Cache` sets.
@@ -118,5 +122,11 @@ keys and values of the reference images of each row of the batch from the
 first step of a run, the later steps of the run reuse those of each row,
 and the end of the run drops them.
 
+`SkipLayerGuidanceDiTSimple` on MOSS-Image is not refused when it names
+`single_layers` alone, with `start_percent` 0 and `end_percent` 1. The
+server leaves those blocks out of the negative pass of each step. The
+node is refused with `double_layers`, another window, a linked widget or
+a block MOSS-Image lacks, and on any other model.
+
 A refusal from a node's own setup shows on that node. A refusal during
 sampling shows on the sampler and names the node upstream that caused it.
diff --git a/docs/comfyui.md b/docs/comfyui.md
@@ -213,6 +213,17 @@ on with a `CFGGuider` and a `KSamplerSelect`. Krea 2 Turbo samples at `cfg`
 Krea 2 quant fits a 1024x1024 image on a Mac with 16 GB, so set the latent
 to 512x512 there.
 
+MOSS-Image reads its prompt with the Qwen3.5-9B GGUF on the server, and
+the gmlx nodes run it on the Mac's GPU without the OpenMOSS ComfyUI plugin
+or the ComfyUI-MOSS-GGUF bridge. Its workflow, `moss-image-t2i`, takes the
+three gmlx loaders with `moss-image` or `moss-image-base` in `model_name`,
+`Empty Flux 2 Latent` for the image's size, `ModelSamplingSD3` at shift 5,
+and `SkipLayerGuidanceDiTSimple` with `single_layers` set to 22, which
+leaves block 22 out of the negative pass, as the OpenMOSS workflow does.
+`KSampler` samples 50 steps at `cfg` 3 with `euler` and `sgm_uniform`.
+That is ComfyUI's schedule, not the release's schedule that
+`gmlx image generate` and the server run.
+
 There are two more gmlx nodes. `gmlx Sigmas` gives the fixed schedule of a
 distilled model such as Qwen-Image 2.1 Turbo, for `SamplerCustom` or
 `SamplerCustomAdvanced`, though `KSampler` works on every model. The
@@ -226,7 +237,7 @@ stock nodes work with a gmlx model and what runs where.
 
 ## The example workflows
 
-ComfyUI's template browser lists ten workflows under `comfyui-gmlx`. Their
+ComfyUI's template browser lists eleven workflows under `comfyui-gmlx`. Their
 files are in `custom_nodes/comfyui-gmlx/example_workflows`. Each has a note
 with the commands that install the gmlx nodes and pull its model.
 
@@ -236,6 +247,7 @@ with the commands that install the gmlx nodes and pull its model.
 | `z-image-turbo-lora-t2i` | The same through a stock `LoraLoader` | A Z-Image LoRA named `z-image-turbo-style.safetensors` in `models/loras`, or another picked in the node |
 | `qwen-image-2.1-edit-two-references` | A Qwen-Image 2.1 edit from two reference images, through `TextEncodeQwenImage21` | `gmlx pull qwen-image`, and `reference-1.png` and `reference-2.png` in ComfyUI's `input` folder |
 | `krea-2-turbo-t2i` | Krea 2 Turbo text to image at 1024x1024 in 8 steps, through `BasicScheduler` and `SamplerCustomAdvanced` | `gmlx pull krea-2-turbo`, and on a Mac with 16 GB a 512x512 latent |
+| `moss-image-t2i` | MOSS-Image text to image at 1024x1024 in 50 steps at `cfg` 3, with block 22 left out of the negative pass through `SkipLayerGuidanceDiTSimple` | `gmlx pull moss-image` |
 | `flux-2-klein-t2i` | FLUX.2 klein text to image at 1024x1024 in 4 steps, through `Flux2Scheduler` and `SamplerCustomAdvanced` | `gmlx pull flux-2-klein` |
 | `flux-2-klein-edit-two-references` | A FLUX.2 klein edit from two reference images, through `ReferenceLatent` | `gmlx pull flux-2-klein`, and `reference-1.png` and `reference-2.png` in ComfyUI's `input` folder |
 | `flux-2-klein-9b-t2i` | FLUX.2 klein 9B text to image at 1024x1024 in 4 steps, the same graph on the 9B model | `gmlx pull flux-2-klein-9b`, and a Mac with 24 GB or more |

diff --git a/docs/image-api.md b/docs/image-api.md
@@ -90,7 +90,7 @@ GIF.
 
 | Model | Edits | Variations |
 |-------|-------|------------|
-| Z-Image, Krea 2 | One image, which the result starts from at `strength`, 0.6 by default. More images get 400. | The image with an empty prompt, at the default `strength` |
+| Z-Image, Krea 2, MOSS-Image | One image, which the result starts from at `strength`, 0.6 by default. More images get 400. | The image with an empty prompt, at the default `strength` |
 | Qwen-Image 2.1 | Up to 10 reference images, when the entry has `encoder_mmproj` | The image as a reference, with an instruction to make a variation of it |
 | FLUX.2 klein | Up to 10 reference images, with no vision tower | The image as a reference, with an instruction to make a variation of it |
 | FLUX.3 klein | Up to 10 reference images, with no vision tower | The image as a reference, with an instruction to make a variation of it |