Qwen3.6-27B-A3B-Coder — code-specialist expert prune

~27B params · 184 experts (72 dropped/layer) · A3B active · top-10 routing (baked) · MTP head + vision tower

A code-targeted expert prune of Qwen3.6-35B-A3B: the mixture is reduced from 256 experts to 184 per layer (72 dropped, ~35B→27B, still A3B active), keeping the code-competent experts. Same router, attention, norms, MTP head and vision tower as the base — only the expert keep-set changes. No fine-tuning, no distillation — pure expert selection.

Full model card & methodology: ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF on Hugging Face.

How it works

The 256-expert teacher is profiled per-expert on a balanced corpus plus targeted LiveCodeBench and MultiPL-E (Rust/Java/JS) pass-response classes. A wmax drop map with the code classes up-weighted evicts 72 experts per layer, protecting the code-competent ones. Serving at top-10 (vs the base top-8) is a routing-recovery dial baked into the config: it buys back instruction-following at no cost to code. Pass --override-kv qwen35moe.expert_used_count=int:8 to any llama.cpp tool to A/B back to native top-8.

Evaluation (Q6_K, llama.cpp, temp 0.6 / top-p 0.95 / top-k 20)

▲ = matches or beats the full 256-expert teacher (AIME, GSM8K, ARC, MultiPL-E). MultiPL-E 0.840 tops the teacher and beats the older LCB-only coder by +17pp; HumanEval 0.970 ties it. Average sits within 0.03 of the full teacher at ~21% fewer experts. The trade is graduate science (GPQA, MATH) and instruction-following (IFEval), not code.

Which tier to pull

Two ladders share this repo. ContribDynamic (CD-*) tiers apply per-layer bit allocation and sit at full-precision code parity — start here. Standard imatrix K-quants cover the same size band. Every tier is built with imatrix (load-bearing at 4-bit on this model). Append vision- to any tag for the multimodal tower (e.g. :vision-CD-IQ4_K_M).

Pull

ollama run mannix/qwen3.6-27b-a3b-coder                   # :latest = Q4_K_M, ~16.5 GB
ollama run mannix/qwen3.6-27b-a3b-coder:CD-IQ4_K_M        # recommended — full-precision-parity code
ollama run mannix/qwen3.6-27b-a3b-coder:vision-CD-IQ4_K_M # + vision tower (multimodal)

Every tag ships these sampling defaults baked in: temperature 1, top_p 0.95, top_k 20, min_p 0, repeat_penalty 1, presence_penalty 1.5, num_ctx 32768. ollama 0.30.x has no reasoning-budget flag, so num_ctx is what bounds the thinking phase there — keep it well above your prompt+output budget.

Hardware note — Blackwell (sm_120)

All ollama tiers in this repo run coherently on every backend, including NVIDIA Blackwell (RTX PRO 6000, sm_120). The two low i-quant tiers hit by a llama.cpp / ggml sm_120 CUDA-kernel bug (IQ2_M, IQ3_M) are intentionally not shipped on ollama — they live on Hugging Face only, for CPU / Ampere / Ada use. Nothing to work around here.

Good to know

• Native MTP head ships in the GGUFs for speculative decoding, alongside the -vision mmproj tower.
• Verbosity on open-ended reasoning (GPQA/AIME) is largely inherited from the base teacher and is bounded by the generation cap; code and boxed-math tasks terminate cleanly.
• imatrix.dat is archived in the GGUF repo — quants are reproducible and auditable.

Links: HF · GGUF + full card · base model

Apache-2.0 · research checkpoint.