models · qwen

Qwen3.6-27B

Qwen3.6 dense 27.8B. 262K context, 150 tok/s. Arena 1138. GPQA Diamond 87.8%, LiveCodeBench 75.8%. Strong dense model at 28B scale.

Compare Verify

What this registry has checked

Quarantine. We hold the publisher’s own published hash for these weights and have recorded it against this entry. We have not downloaded and hashed the weights ourselves, so the hash is the publisher’s claim, not our measurement.

Weight file SHA-256	20e9d0e734a8fb4e4f9a2ddd9aead33317e737c2b751bc601b6bec0173ab5237
Method	hf-published-sha256

Re-check this yourself → Source → Full record →

What is recorded

Publisher	qwen
Origin	Recorded as a covered nation under 10 U.S.C. § 4872(f) — the PRC, Russia, Iran or North Korea. The registry did not record which.
Licence	Apache-2.0
Parameters	27,781,427,952 (27.8B)
Context window	262K
Modality	text+image+video->text
Native precision	BF16
Pulls recorded here	420,014
Downloads reported upstream	4,048,692

Reported scores

GPQA Diamond	84.2
IFBench	67.6
Humanity's Last Exam	23.1

Reported by a third party and recorded here, not re-run by us. Source: artificialanalysis.ai/api/v2.

Measured on real hardware

None of these runs are ours. Each figure is the reporter’s, linked to their own evidence, under their own trust label. The number shown is single-stream decode — one user, concurrency 1 — because aggregate throughput across many users is a different quantity and reads as far faster than anyone will see.

Reported to local-ai-registry

Hardware	Engine	Build measured	Decode, 1 stream	Label	Reported by
Radeon AI PRO R9700	hipfire	MQ4	278 tok/s @ 2,048	candidate	localmaxxing
Radeon RX 7900 XTX	hipfire	MQ4	252 tok/s @ 2,048	candidate	localmaxxing
GeForce RTX 3090 Ti	llama.cpp	UD-Q4_K_XL	156 tok/s @ 4,096	candidate	localmaxxing
GeForce RTX 3090	llama.cpp	UD-Q4_K_XL	150 tok/s @ 2,048	candidate	localmaxxing
RTX PRO 6000 Blackwell	llama.cpp	Q4_0	144 tok/s @ 4,445	candidate	localmaxxing
RTX PRO 6000 Blackwell	vllm	NVFP4	140 tok/s @ 2,048	candidate	localmaxxing
Ryzen AI Max+ 395	hipfire	MQ4	112 tok/s @ 2,048	candidate	localmaxxing
GeForce RTX 5090 ×2 devices	vllm	safetensors NVFP4	85.3 tok/s @ 8,192	candidate	localmaxxing
GeForce RTX 5090	llama.cpp	safetensors IQ4_XS	79.2 tok/s @ 8,192	candidate	localmaxxing
GeForce RTX 3090 ×2 devices	vllm	safetensors FP8	77.8 tok/s @ 8,192	candidate	localmaxxing

From local-ai-registry (MIT), commit 124959d: 306 runs across 32 machines, 10 validated — model revision and runtime pinned, launch accepted — and 296 candidate, their label for useful evidence without a reproducible-launch promise. 198 more not listed. 3 report figures that contradict each other — prefill below decode, which parallel prefill essentially never is — and are left to the source rather than ranked here.

Measured by local.ai

Hardware	Build measured	Decode @ 8K	Memory @ 8K	Decode mode
NVIDIA RTX PRO 6000 Blackwell 96GB	Q4_K_M	74.4 tok/s	19.5 GB	ordinary
NVIDIA GeForce RTX 5090 32GB	NVFP4	73.9 tok/s	28.2 GB	ordinary
NVIDIA RTX PRO 6000 Blackwell 96GB	NVFP4	73.0 tok/s	90.2 GB	ordinary
NVIDIA GeForce RTX 5090 32GB	NVFP4	65.8 tok/s	28.3 GB	ordinary
NVIDIA RTX PRO 6000 Blackwell 96GB	NVFP4	64.7 tok/s	92.0 GB	ordinary
NVIDIA RTX 6000 Ada 48GB	Q4_K_M	44.5 tok/s	22.7 GB	ordinary
NVIDIA RTX 6000 Ada 48GB	NVFP4	38.0 tok/s	43.4 GB	ordinary
Mac Studio M3 Ultra 96GB · 60-core GPU	4bit	28.9 tok/s	23.2 GB	ordinary

Measured by local.ai, not by us: 112 runs, 104 more not listed. Each names its engine pinned by image digest and the exact serve command, in the full record.

Elsewhere in the registry

All models → Browse with filters → What will run on my hardware? →

Registry record as of 2026-06-11. Identity, licence and parameter count are properties of a release and do not move; tier can, and the verify link above re-checks it against the live registry rather than this page.