---

license: apache-2.0

language:

- en

- zh

- multilingual

tags:

- gguf

- llama.cpp

- qwen

- qwen3.6

- moe

- multimodal

- image-text-to-text

- mxfp4

- mtp

- speculative-decoding

pipeline_tag: image-text-to-text

base_model:

- HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive

- khoi11/Qwen3.6-35B-A3B-Uncensored-HauhauCS-FP16

---

hauhau_qwen36_35b_a3b_mxfp4_gguf_mtp_visualsupport

This repo contains the local serving package for HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive using a robinshao-style MXFP4_MOE main GGUF, a separate MTP draft GGUF for speculative decoding, and a separate multimodal mmproj GGUF for image input support in patched llama.cpp.

Files

• hauhau_qwen36_35b_a3b_mxfp4_robinstyle.gguf: main text model in robinshao-style MXFP4 layout
• hauhau_qwen36_35b_a3b_mxfp4_mtp.gguf: MTP draft GGUF used with --spec-type draft-mtp
• mmproj-Hauhau_Qwen36_35B_A3B_Fp16-Q8_0.gguf: multimodal projector / vision encoder GGUF
• llama.cpp.src-patched.zip: patched local llama.cpp source snapshot used for conversion and serving
• build_result.json: build and validation summary
• mtp-benchmark.json: short local speed comparison for the text+MTP serving pair

Package Notes

• Source FP16 weights: khoi11/Qwen3.6-35B-A3B-Uncensored-HauhauCS-FP16
• Upstream model lineage: HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive
• Main model and draft model reuse the previously validated robinshao-style MXFP4 + MTP export
• Visual support is provided by a newly exported mmproj GGUF from the original multimodal FP16 checkpoint
• This serving package is multimodal only when the main model is loaded together with the mmproj file

Local Deployment Notes

Validated locally on Windows with patched llama.cpp using speculative decoding and multimodal input:

llama-server -m hauhau_qwen36_35b_a3b_mxfp4_robinstyle.gguf   --mmproj mmproj-Hauhau_Qwen36_35B_A3B_Fp16-Q8_0.gguf   --spec-type draft-mtp   --spec-draft-model hauhau_qwen36_35b_a3b_mxfp4_mtp.gguf   --spec-draft-device CUDA0   --spec-draft-ngl all   --jinja   --device CUDA0   --gpu-layers all   --ctx-size 8192   --batch-size 512   --ubatch-size 128   --flash-attn on   --cache-type-k f16   --cache-type-v f16   --cache-ram 0

Local validation confirmed:

• the model still loads with draft-mtp
• the mmproj loads successfully
• /v1/chat/completions accepts image input
• a real image smoke test identified the primary object correctly

Caveats

• The main GGUF and MTP draft are still the same text-focused MXFP4 exports; visual support comes from the additional mmproj file.
• If you omit --mmproj, the package behaves like the earlier text-only MTP release.
• Native training context metadata remains 262144, but local no-offload limits still depend on VRAM, KV cache settings, and whether the draft model is loaded.