Qwen3.6-27B

Alibaba Qwen · Released: unverified (Ollama readme says it followed the Feb Qwen3.5 release; GitHub issues about it date from 2026-04-28) · Last verified 2026-10-16

Dense 27B hybrid-thinking model, popular for local agentic coding. Most pain is chat-template / tool-call handling in agent loops.

Recommended quants

Unsloth Dynamic GGUF UD-Q4_K_XL (~18 GB total memory at 4-bit)
unsloth/Qwen3.6-27B-GGUF
Unsloth's default for its llama.cpp guide.
Unsloth MTP GGUF
unsloth/Qwen3.6-27B-MTP-GGUF
Speculative decoding; llama.cpp merged MTP support per Unsloth. Unsloth suggests --spec-draft-n-max 2.
Unsloth NVFP4
unsloth/Qwen3.6-27B-NVFP4
Blackwell only (RTX 50, DGX Spark). vLLM.
MLX 3/4/6/8-bit, MXFP4, NVFP4
Unsloth MLX uploads; Ollama tag qwen3.6:27b-mlx
Apple Silicon.

Chat template Caveats

Official template has known tool-call issues in agent loops (empty tool calls, reasoning leaking into tool turns). Community fixed templates exist.

Fix: Use a community fixed template, e.g. froggeric/Qwen-Fixed-Chat-Templates (llama.cpp: --jinja --chat-template-file chat_template.jinja --reasoning-format deepseek). The StableQuant rebuild project is on hold and points to spiritbuun/buun-Qwen3.6-chat_template.

Sampling settings

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`
Disable thinking	`--chat-template-kwargs '{"enable_thinking":false}'`

Runtimes

vLLM	Caveats	Runs (Unsloth NVFP4 guide). Empty-tool-call issue reported on vLLM 0.19.1 and 0.20.0 in issue #150.
llama.cpp	Works	Documented by Unsloth incl. MTP. Do not use CUDA 13.2 (gibberish per Unsloth).
Ollama	Works	Official tags qwen3.6:27b and qwen3.6:27b-mlx.
LM Studio	Unverified	froggeric template lists LM Studio as supported; model behavior not verified here.
MLX	Works	Unsloth MLX quants + Ollama MLX tag.

Tool calling in agent tools

Claude Code (local endpoint)	Caveats	Ollama documents `ollama launch claude --model qwen3.6`. Issue #150: model stops with empty tool call under Claude Code + vLLM.
OpenCode	Caveats	Ollama documents `ollama launch opencode --model qwen3.6`. OpenCode bug shows thinking as plain text; workaround --reasoning-format deepseek (StableQuant notes).
Cline	Unverified
Aider	Unverified

Known issues

Sources checked

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