trustos/Qwen3.6-35B-A3B-MTPLX-Vision-MTP

Qwen3.6-35B-A3B · MTPLX · Vision + MTP (unified coder + VLM)

One local model that both codes at Qwen3.6-35B level and sees images, for Apple Silicon via MTPLX native-MTP serving.

What this is

Qwen/Qwen3.6-35B-A3B ships multimodal (Qwen3_5MoeForConditionalGeneration, 333 model.visual.* vision-tower tensors) — but common MLX/MTPLX text conversions strip the vision tower, leaving a text-only build with dangling image_token_ids. This repo restores it:

• Body: MTPLX 8-bit forge of Qwen3.6-35B-A3B (affine, group_size=64) — an MoE tuned for fast agentic coding on Apple Silicon.
• MTP sidecar (mtp.safetensors): native multi-token-prediction draft head → MTPLX speculative decoding (~100 tok/s decode ladder on an M-series Max).
• Vision tower (vision_tower.safetensors, 333 tensors, bf16 unquantized): extracted from the official checkpoint's own model.visual.* weights (shards 1–2), renamed to the vision_tower.* prefix. Because the tower comes from the same checkpoint as the body, vision-language alignment is preserved by construction — no fine-tuning, no cross-model transplant.
• Official multimodal chat_template.jinja included.

Measured (M5 Max 128 GB, MTPLX 2.0.1)

Test	Result
Coding: 429-line real-module repair (gated agentic loop)	converged 1 turn / 83.6 s — parity with the text-only build (80 s)
Coding: small py/ts bug fixtures	1 turn, 2.6 s / 2.7 s
OpenAI-style tools calls (4 shapes incl. no-tool-needed)	4/4, identical to the text-only reference
Vision: synthetic shape/color/text reads	2/2 exact (incl. text color)
MTP speculative decode warm-up ladder	~97–103 tok/s

Caveats: single-trial agentic benchmarks on one machine; vision checks are synthetic smoke tests, not a VLM benchmark suite.

Run it (MTPLX)

mtplx quickstart --model <this-repo-dir> --port 8092 \
  --chat-template-path <this-repo-dir>/chat_template.jinja
# OpenAI-compatible: POST /v1/chat/completions with text and/or image_url content parts

Reasoning ON by default; serve with --reasoning off for terse agentic use.

Provenance & license

Derived solely from Qwen/Qwen3.6-35B-A3B (Apache-2.0). Quantization + MTP packaging via MTPLX forge; vision tower restored verbatim (bf16) from the same checkpoint. No third-party weights. Apache-2.0, same as the base.