Qwen3.6-35B-A3B

Alibaba Qwen · Released: unverified · Last verified 2026-10-16

MoE sibling of Qwen3.6-27B (3B active). Faster decode; same template family and same tool-call caveats.

Recommended quants

Unsloth Dynamic GGUF UD-Q4_K_XL (~23 GB at 4-bit)
unsloth/Qwen3.6-35B-A3B-GGUF
Unsloth says its GGUFs top KLD in 21 of 22 sizes (vendor claim).
Unsloth MTP GGUF
unsloth/Qwen3.6-35B-A3B-MTP-GGUF
MoE gains ~1.15-1.25x per Unsloth.
Unsloth NVFP4 / NVFP4-Fast
unsloth/Qwen3.6-35B-A3B-NVFP4(-Fast)
Blackwell only. On DGX Spark use --moe-backend flashinfer_b12x or expect ~2x slower (Unsloth).
MLX 3/4/8-bit
Unsloth MLX; Ollama qwen3.6:35b-mlx

Shares the Qwen3.6 template issues. A multi-turn system-message failure on 35B-A3B is referenced from issue #150 (issue #164, not opened here).

Fix: Same community templates as 27B (froggeric v22.5 covers all Qwen 3.5/3.6/3.8 sizes).

Thinking, general	`temp 1.0, top_p 0.95, top_k 20, min_p 0, presence_penalty 0`
Thinking, precise coding	`temp 0.6, top_p 0.95, top_k 20, min_p 0, presence_penalty 0`
Non-thinking	`temp 0.7, top_p 0.8, top_k 20, min_p 0, presence_penalty 1.5`

vLLM	Caveats	Runs (Unsloth). Do not force Marlin backend (2.5x slower per Unsloth). A user in #150 reports the empty-tool-call issue on 35B-A3B via vLLM 0.20.2.
llama.cpp	Works	Documented by Unsloth incl. MTP.
Ollama	Works	qwen3.6:35b (also :latest) and qwen3.6:35b-mlx.
LM Studio	Unverified
MLX	Works	Unsloth MLX + Ollama MLX tag.

Claude Code (local endpoint)	Unverified	Ollama documents the launch path; reliability not verified for this size.
OpenCode	Unverified	Ollama documents the launch path.
Cline	Unverified
Aider	Unverified

Spotted something wrong or a new fix? Corrections are reviewed before publishing. (Submission form coming soon.)