feat(k2-horizon): K2 Horizon 7B and MoVA 36B-A4B

CHANGELOG.md

@@ -23,6 +23,8 @@ adhere to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
   and image and audio input through its `--mmproj` companion.
 - An image LoRA whose PEFT header sets `use_rslora` loads, scaled as PEFT
   scales it, where it was refused.
+- K2 Horizon 7B and K2 Horizon MoVA 36B-A4B (GGUF arch `k2-horizon`), with
+  `reasoning-low`, `-medium` and `-high` intents and tool calls.
 
 ### Changed
 
@@ -39,6 +41,10 @@ adhere to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ### Fixed
 
+- A chat template with a `{% generation %}` block was refused by
+  `--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 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

README.md

@@ -139,6 +139,7 @@ back.
 - GLM 4 through 5.3, Kimi-K3, MiniMax M2 and M3, and gpt-oss
 - Hunyuan, Hy3, HY4, Muse Glimmer and Muse Spark
 - Granite, Nemotron-H and Falcon-H1
+- K2 Horizon, dense and MoVA
 
 Each family's output is checked against llama.cpp at 16k context. The
 [coverage table](https://asher.github.io/gmlx/arch-coverage.html) lists the

THIRD_PARTY_NOTICES.md

@@ -10,7 +10,7 @@ in [`licenses/`](licenses/).
 | [mlx-lm](https://github.com/ml-explore/mlx-lm) | MIT | Release 0.31.3 vendored whole at `gmlx/_vendor/mlx_lm`, with its LICENSE, its module references rewritten and each gmlx edit marked `# gmlx:`. Model/cache classes from unmerged PRs: `gmlx/models/deepseek_v4/model.py` + `gmlx/models/deepseek_v4/hyper_connection.py` + `gmlx/models/deepseek_v4/cache.py` (PR #1192), `gmlx/models/minimax_m3.py` (PR #1401), `gmlx/models/hy_v3/model.py` (PR #1485); `gmlx/models/kimi_k3.py` builds on the released `kimi_linear` module. The KV cache classes in `gmlx/cache/kv_cache.py` and the mask and SDPA helpers in `gmlx/cache/attention.py` are mlx-lm's `models/cache.py` and `models/base.py` code, by way of mlx-vlm 0.7.3. The FunctionGemma tool parser in `gmlx/models/function_gemma.py` comes from release 0.31.3, by way of mlx-vlm 0.6.15. The Gemma 4 text weight sanitize in `gmlx/models/gemma4/lm_sanitize.py` comes from release 0.31.3. |
 | [mlx-vlm](https://github.com/Blaizzy/mlx-vlm) | MIT; Apache-2.0 for eight files | Release 0.7.3 vendored whole at `gmlx/_vendor/mlx_vlm`, with its LICENSE, its module references rewritten and each gmlx edit marked `# gmlx:`. Eight of its files keep an Apache-2.0 header, under `models/llada_image`, `models/mage_vl`, `models/nemotron_voicechat` and `models/sam3d_objects`. Also the qwen3.5 and qwen3.5 MoE text models of release 0.6.15, vendored in `gmlx/models/qwen35/` (`language.py`, `gated_delta.py`, `moe_language.py`); the Gemma 4 text model and unified container of release 0.7.3, vendored in `gmlx/models/gemma4/` (`language.py`, `unified.py`); the KV cache classes of release 0.7.3 (`models/cache.py`), vendored whole in `gmlx/cache/kv_cache.py`, and its mask and SDPA helpers (`models/base.py`) in `gmlx/cache/attention.py`; the chunked-eval `store_kv_blocks` override in `gmlx/cache/apc_manager.py`, the thinking-budget logits processor in `gmlx/gen/thinking_budget.py`, and `entropy_transfer_mask` in `gmlx/systemone/denoise.py`. |
 | [omlx](https://github.com/jundot/omlx) | Apache-2.0 | DeepSeek-V4-Flash MTP port (`gmlx/models/deepseek_v4/mtp.py`), the MTP undo log of the rotating cache classes in `gmlx/cache/kv_cache.py`, and the native-kernel dispatch pattern in `gmlx/models/deepseek_v4/model.py`. |
-| [llama.cpp / ggml](https://github.com/ggml-org/llama.cpp) | MIT | Pre-tokenizer split regex patterns in `gmlx/load/tokenizer.py`; GGUF metadata and quantization-type conventions; the Muse Glimmer vision tower in `gmlx/models/muse_glimmer/vlm_model.py` (from `clip_graph_muse_glimmer` and `tools/mtmd/clip.cpp`) and its projector constants in `gmlx/load/vlm.py`; the Muse Spark audio encoder in `gmlx/models/muse_spark/audio.py` (from `clip_graph_muse_spark_audio`); the Kimi-K3 chat-template fixture `tests/fixtures/kimi_k3_template.jinja` (PR #26185). |
+| [llama.cpp / ggml](https://github.com/ggml-org/llama.cpp) | MIT | Pre-tokenizer split regex patterns in `gmlx/load/tokenizer.py`; GGUF metadata and quantization-type conventions; the Muse Glimmer vision tower in `gmlx/models/muse_glimmer/vlm_model.py` (from `clip_graph_muse_glimmer` and `tools/mtmd/clip.cpp`) and its projector constants in `gmlx/load/vlm.py`; the Muse Spark audio encoder in `gmlx/models/muse_spark/audio.py` (from `clip_graph_muse_spark_audio`); the Kimi-K3 chat-template fixture `tests/fixtures/kimi_k3_template.jinja` (PR #26185); the K2 Horizon chat-template fixture `tests/fixtures/k2_horizon_template.jinja` (`models/templates/IFM-K2-Horizon.jinja`). |
 | [ds4.c (dwarfstar)](licenses/ds4.c-LICENSE) | MIT | QAT round-trip and sink-attention reference semantics reimplemented in `gmlx/models/deepseek_v4/model.py` (cited per function in comments); the DSpark drafter runtime reimplemented in `gmlx/models/deepseek_v4/dspark.py`; the `deepseek4-dspark` sidecar layout (`gguf-tools/deepseek4-quantize.c` and the published DSpark-support sidecar) emitted by `scripts/convert_dspark_sidecar.py`. |
 | [DeepSeek-V4-Flash-0731](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731) | MIT | Reference DSparkBlock forward semantics (`inference/model.py`) reimplemented in `gmlx/models/deepseek_v4/dspark.py` (license text in [`licenses/deepseek-v4-LICENSE`](licenses/deepseek-v4-LICENSE)). |
 | [beellama.cpp](https://github.com/Anbeeld/beellama.cpp) | MIT | The KVarN KV-cache record format and cache semantics (128-token rotated records with three axis vectors, fp16 sink and precision tail, `kvarnN` widths, `--kv-tail-tokens`) reimplemented in `gmlx/cache/kvarn_cache.py` and `gmlx/cache/kvarn_sdpa.py`. The method is Huawei's KVarN ([arXiv:2606.03458](https://arxiv.org/abs/2606.03458), reference at [huawei-csl/KVarN](https://github.com/huawei-csl/KVarN), Apache-2.0; no code from it is used). The kernels live in mlx-kquant, whose notices cover its own beellama.cpp port. |

docs/arch-coverage.md

@@ -1,6 +1,6 @@
 # Supported architectures
 
-gmlx maps 44 GGUF architectures to model types, and 43 of them load end-to-end
+gmlx maps 45 GGUF architectures to model types, and 44 of them load end-to-end
 with no `hf_source`. `scripts/check-coverage.py` generates this page from the
 architecture table and the mlx-lm that gmlx carries, so do not edit it by hand.
 
@@ -47,6 +47,7 @@ family, backend, status and caveat:
 | `hunyuan-moe` | hunyuan | mlx-lm | loadable |  |
 | `hy_v3` | hunyuan | mlx-lm | loadable | Early GGUFs whose arch is hy-v3 with a dash are not mapped. Reconvert them. |
 | `hyv4` | hunyuan | mlx-lm | loadable | The GGUF has no MTP head, so speculative decoding is not available. `--kv-bits` is refused. |
+| `k2-horizon` | k2-horizon | mlx-lm | loadable | The GGUF has no MTP head, so speculative decoding is not available. |
 | `kimi-k3` | kimi | mlx-lm | loadable | The hybrid cache cannot go back in place, so after `/retry` or `/undo`, chat reads the earlier conversation again with the next message. |
 | `llama` | llama | mlx-lm | loadable |  |
 | `minimax-m2` | minimax | mlx-lm | loadable |  |

docs/family-defaults.md

@@ -39,6 +39,7 @@ a model ends up with.
 | `kimi-k2` | `deepseek2 named (?i)\bkimi` | temperature=1.0 top_p=0.95 | - |
 | `muse` | `muse-glimmer` | temperature=1.0 top_p=0.95 top_k=64 thinking_start_token=<\|start\|>assistant to=self<\|message\|> thinking_end_token=<\|eom\|> | `@reasoning-high`: reasoning_strength=high; `@reasoning-low`: reasoning_strength=low; `@reasoning-medium`: reasoning_strength=medium; `@reasoning-xhigh`: reasoning_strength=xhigh |
 | `muse-spark` | `muse-spark` | temperature=1.0 top_p=1.0 top_k=64 thinking_start_token=<\|start\|>assistant to=self<\|message\|> thinking_end_token=<\|eom\|> | `@reasoning-high`: reasoning_effort=high; `@reasoning-low`: reasoning_effort=low; `@reasoning-medium`: reasoning_effort=medium; `@reasoning-xhigh`: reasoning_effort=xhigh |
+| `k2-horizon` | `k2-horizon` | temperature=1.0 top_p=0.95 | `@reasoning-high`: reasoning_effort=high; `@reasoning-low`: reasoning_effort=low; `@reasoning-medium`: reasoning_effort=medium |
 | `llama` | `llama`, `smollm3` | temperature=0.6 top_p=0.9 | - |
 | `mistral` | `mistral3` | temperature=0.15 | - |
 | `default` | (anything else) | temperature=0.7 top_p=0.95 | `@coding`: temperature=0.3; `@creative`: temperature=1.0 min_p=0.05 |

docs/long-context.md

@@ -79,6 +79,7 @@ flags refuses.
 | `gpt-oss` | Sliding window layers keep the header's factor 32. |
 | `granite`, `granitehybrid` | |
 | `hy_v3`, `hyv4` | |
+| `k2-horizon` | |
 | `llama` | |
 | `minimax-m2`, `minimax-m3` | |
 | `mistral3` | |

gmlx/_vendor/manifest.json

@@ -714,10 +714,10 @@
   "mlx_vlm/models/jina_vlm/language.py": {"status": "verbatim", "upstream": "65f112e5c77191c8e29cb01991ae21b5cf9d693cbf651664e27adbd25727b63c"},
   "mlx_vlm/models/jina_vlm/processing_jinavlm.py": {"status": "verbatim", "upstream": "81fdae0d704f82f00081cb9c2dfd3ff4fe0f680f3275251eeab63ce6e946d875"},
   "mlx_vlm/models/jina_vlm/vision.py": {"status": "verbatim", "upstream": "424382f0520c3ee7a0f6482c48a60f3d8498fc5567da39e395d87b42b88cfbf5"},
-  "mlx_vlm/models/k2_horizon/__init__.py": {"status": "verbatim", "upstream": "ab9041063ffb321cc77fa4b03f2ef24c074af5de7d4f5ae832d05674a8da9081"},
-  "mlx_vlm/models/k2_horizon/config.py": {"status": "verbatim", "upstream": "d2f6c6254fd7072612c1423c72846cd6c4f857ee5ab51eb3988438281b370ecb"},
-  "mlx_vlm/models/k2_horizon/k2_horizon.py": {"status": "verbatim", "upstream": "ff47fd80ceb15716553a154029853abe080c5569d3fa8c9406d1bc85dbf8a180"},
-  "mlx_vlm/models/k2_horizon/language.py": {"status": "verbatim", "upstream": "9b0df74e8226b437ad242fe488c8881737c4e75ee534b192136fd596cc63032c"},
+  "mlx_vlm/models/k2_horizon/__init__.py": {"status": "excluded", "upstream": "ab9041063ffb321cc77fa4b03f2ef24c074af5de7d4f5ae832d05674a8da9081"},
+  "mlx_vlm/models/k2_horizon/config.py": {"status": "excluded", "upstream": "d2f6c6254fd7072612c1423c72846cd6c4f857ee5ab51eb3988438281b370ecb"},
+  "mlx_vlm/models/k2_horizon/k2_horizon.py": {"status": "excluded", "upstream": "ff47fd80ceb15716553a154029853abe080c5569d3fa8c9406d1bc85dbf8a180"},
+  "mlx_vlm/models/k2_horizon/language.py": {"status": "excluded", "upstream": "9b0df74e8226b437ad242fe488c8881737c4e75ee534b192136fd596cc63032c"},
   "mlx_vlm/models/kernels.py": {"status": "verbatim", "upstream": "7f18031e867b31eae7f57f38c53c653983bcd867e65080e66a32f60da8452360"},
   "mlx_vlm/models/kimi_k25/__init__.py": {"status": "verbatim", "upstream": "647b1ea761370bbf2c7f0de5ca500bd185e837994abd04246035bb689b3f3e98"},
   "mlx_vlm/models/kimi_k25/config.py": {"status": "verbatim", "upstream": "784f6bddddbaf0d38135e09e2596b5856573aafa551d3d628f246a5428e20034"},
@@ -1525,7 +1525,7 @@
   "mlx_vlm/server/realtime.py": {"sha256": "942671328088936df3145eb72ba7e6650d0171c22ca17321d79ad113cd0482b9", "status": "codemod", "upstream": "815998c835bd348649779ce480152c48cf8ff770ca53503f1a1c31649668b416"},
   "mlx_vlm/server/request_normalization.py": {"sha256": "38c69ac6dd59534f5b6ad93adf15be145bf299c1aae9af0c9d0216338c006c59", "status": "edited", "upstream": "37eb0727a61eb8a647cf010af20098f388688336d0e9df9f217c2409fc3a00e5"},
   "mlx_vlm/server/reranking.py": {"status": "verbatim", "upstream": "36b03fb077741261696795a3b32bfcae64339d0d50d091ee39f2df4b97343179"},
-  "mlx_vlm/server/responses_state.py": {"sha256": "58919f35ec6bed9a40390ce45e9af28d65dbdf52e166ace2854afd9267167862", "status": "codemod", "upstream": "138184cb972c43c6f48fc104642608add2605180c1682d3cc1dfb0bd408f3657"},
+  "mlx_vlm/server/responses_state.py": {"sha256": "0a195e66a156624c7fc8710c54fb3a782f6a14f7bfa2cd75b59e099e060eebf4", "status": "edited", "upstream": "138184cb972c43c6f48fc104642608add2605180c1682d3cc1dfb0bd408f3657"},
   "mlx_vlm/server/runtime.py": {"status": "verbatim", "upstream": "de80f06d985f98270d7f5d16d2c4e41f1f8ecd69c15c9a819ec8e891ff1f83d2"},
   "mlx_vlm/server/runtime_config.py": {"sha256": "ae05cb5e399e51700340a7ed31e24f7385f5373325e37fe15425ceee5e570b8a", "status": "codemod", "upstream": "cccded87d36ee60bb302be4898b2a674335448e9eb6838584ca797095e1b4c07"},
   "mlx_vlm/server/schemas.py": {"status": "verbatim", "upstream": "80c863e32dd543ca878876e3b183559adfa1a27848d44599ae0852efe74fea68"},