license: apache-2.0 library_name: mlx-vlm base_model: huihui-ai/Huihui-ThinkingCap-Qwen3.6-27B-abliterated tags:
- mlx
- direct-mlx-vlm
- mlx-vlm
- qwen3_5_mtp
- qwen3.6
- mtp
- speculative-decoding
- draft-model
- bf16
Huihui ThinkingCap Qwen3.6 27B — direct mlx-vlm MTP drafter (BF16)
This is a standalone MTP draft model for direct mlx-vlm use. It is not a standalone chat model.
Current LM Studio limitation: despite the historical repository name, LM Studio
0.4.19+2with MLX runtime1.10.1does not support speculative draft models for batched vision models. Do not select this as an LM Studio primary model or expect LM Studio to attach it to this Qwen3.6 VLM. For MTP, use the oMLX Native-MTP model, which is the recommended faster and more mature integrated path.
Use it with this target model:
Direct mlx-vlm
uv run mlx_vlm.generate \
--model pixelkaiser/Huihui-ThinkingCap-Qwen3.6-27B-abliterated-MLX-4bit \
--draft-model pixelkaiser/Huihui-ThinkingCap-Qwen3.6-27B-abliterated-MLX-LMStudio-MTP-Drafter-BF16 \
--draft-kind mtp \
--draft-block-size 3 \
--prompt "Hello" \
--max-tokens 128
Compatibility
| Runtime | Use this repository as |
|---|---|
Direct mlx-vlm | MTP draft model, paired with the linked 4-bit target |
| LM Studio MLX | MTP unsupported for this batched VLM in runtime 1.10.1; use the target-only model without MTP |
| oMLX | Not this artifact; use the oMLX indexed-MTP target |
| MTPLX | Not this artifact; use the MTPLX sidecar target |
These runtimes currently use different MTP packaging contracts; the artifacts are intentionally separate.
Verified runtime
Verified end-to-end on Apple Silicon by invoking the mlx-vlm 0.6.3 and MLX 0.31.2 packages bundled with LM Studio's MLX backend directly. A target-plus-drafter smoke generated successfully with speculative decoding active (100% of drafted tokens accepted in the bounded test; acceptance depends on the prompt). This verifies the drafter and mlx-vlm, not LM Studio's batched_vision wrapper.
Technical details
- MTP model type:
qwen3_5_mtp - Block size:
3 - MTP tensors: 15
- MTP precision: BF16
- Target trunk: MLX affine 4-bit, group size 64
- Source revision:
44f63da8141407af529405c1e4b83fa39b70abe0
The drafter was extracted without changing tensor values from the MTP weights preserved in the exact upstream-derived model.
Upstream
- Model: huihui-ai/Huihui-ThinkingCap-Qwen3.6-27B-abliterated
- Base architecture: Qwen/Qwen3.6-27B
- Runtime: LM Studio /
mlx-vlm