Qwen3.8-27B

Alibaba Qwen · Released: unverified (Unsloth GGUF update dated Aug 19; froggeric added 3.8 support 2026-08-13) · Last verified 2026-10-16

Newest dense Qwen with reasoning_effort control. Biggest gotcha: official template defaults to xhigh reasoning, which can burn the token budget.

Recommended quants

Unsloth Dynamic V3.0 GGUF UD-Q4_K_XL (16-19 GB at 4-bit)
unsloth/Qwen3.8-27B-GGUF
MTP enabled; ~1-2 GB extra headroom for MTP.
Unsloth NVFP4
unsloth/Qwen3.8-27B-NVFP4
Blackwell; vLLM, or SGLang v0.5.19+.

Chat template Caveats

Official 3.8 template defaults reasoning to xhigh (can return empty content), injects duplicate blank think blocks in history.

Fix: froggeric/Qwen-Fixed-Chat-Templates v22.x defaults to medium and fixes the history bug. Or set --chat-template-kwargs '{"reasoning_effort":"medium"}' (Unsloth).

Sampling settings

Thinking	`temp 1.0, top_p 0.95, top_k 20, min_p 0, presence_penalty 0, repetition_penalty 1.0`
Non-thinking	`temp 0.7, top_p 0.80, top_k 20, min_p 0, presence_penalty 1.5`
Reasoning effort	`xhigh (default) / medium / low / none via reasoning_effort`

Runtimes

vLLM	Works	NVFP4 documented by Unsloth. froggeric: use --reasoning-parser qwen3 --tool-call-parser qwen3_xml.
llama.cpp	Works	Documented by Unsloth.
Ollama	Unverified	Not checked.
LM Studio	Unverified	froggeric template supports LM Studio; model not verified.
MLX	Unverified	froggeric lists MLX/oMLX template support; quants not verified.

Tool calling in agent tools

Claude Code (local endpoint)	Unverified	froggeric maps Claude Code effort aliases; behavior not verified.
OpenCode	Unverified
Cline	Unverified
Aider	Unverified

Known issues

Sources checked

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