Model

Qwen3.6-27B

OPEN

Alibaba · Qwen3.6 family · released Apr 22, 2026

Not in the original flagged leads, but discovered via direct primary-source research (qwen.ai/blog and github.com/QwenLM/Qwen3.6) while verifying the flagged models -- and arguably the single most important model actually missing, since it is explicitly billed by Alibaba as surpassing the prior open-weight flagship Qwen3.5-397B-A17B (15x its parameter count) on every major agentic-coding benchmark despite being fully dense at 27B params.

ReasoningCodingVisionFunction callingTool useAgentic

2027.4

Elo · rank #76

Parameters

27B

Active params

Undisclosed

Context

256K tokens

Architecture

Hybrid-attention dense model (Qwen3-Next lineage, no MoE routing): 64 layers, 16x(3x(Gated DeltaNet -> FFN) -> 1x(Gated Attention -> FFN)); hidden dim 5120, FFN intermediate dim 17408; natively multimodal (Causal LM with built-in vision encoder)

License

Apache 2.0

Languages

—

API price (in/out)

$0.6 / $3.6

Modalities

text · vision

Benchmark results

Bar shows position within the tracked field; marker = field best

AIMEMath94.1%#23

best: GPT-5.2 · 100.0%

AndroidWorldAgents70.3%#4

best: Qwen3.8-Flash-Next · 84.5%

C-EvalKnowledge91.4%#11

best: Qwen3.6-Plus · 93.3%

CharXivVision78.4%#20

best: Qwen3.8-Flash-Next · 90.6%

DeepSearchQAAgents71.1%#6

best: GPT-5.6 Sol · 93.0%

DeepSWECoding13.3%#16

best: Muse Spark 1.3 · 75.4%

GAIA2Agents40.0%#2

best: Muse Glimmer 30B · 43.3%

GPQA DiamondReasoning87.8%#45

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning24.0%#65

best: Claude Opus 5 · 64.7%

JobBenchAgents21.8%#9

best: Claude Opus 5 · 65.7%

LiveCodeBenchCoding83.9%#31

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MathVistaVision87.4%#5

best: Seed 2.1 Pro · 90.7%

MMBench (English)Vision92.3%#5

best: Qwen3.5-397B-A17B · 93.7%

MMLU-ProKnowledge86.2%#22

best: Claude Fable 5 · 91.5%

MMLU-ReduxKnowledge93.5%#11

best: Qwen3.7-Max · 95.0%

MMLUKnowledge84.5%#63

best: OpenAI o3 · 92.9%

MMMU-ProVision75.8%#27

best: Claude Opus 4.7 · 85.5%

MMMUVision82.9%#16

best: Claude Fable 5 · 89.3%

NL2RepoCoding36.2%#6

best: GLM-5.3 · 58.0%

OCRBenchVision89#17

best: InternVL3-78B · 906

OSWorld-VerifiedAgents63.9%#21

best: Claude Fable 5 · 85.0%

RefCOCOVision92.5%#8

best: Qwen3.5-Omni-Plus · 95.0%

SciCodeCoding39.8%#2

best: Muse Glimmer 30B · 43.6%

ScreenSpot-ProVision76.1%#4

best: GPT-6 Astra · 92.7%

SuperGPQAReasoning66.0%#9

best: Qwen3.7-Max · 73.6%

SWE-bench ProCoding53.5%#41

best: Claude Fable 5.1 · 81.2%

SWE-bench VerifiedCoding77.2%#34

best: Claude Opus 5 · 96.0%

Terminal-Bench 2.0Coding59.3%#46

best: Gemini 3.8 Flash · 89.4%

Video-MMEVision87.7%#5

best: Seed 2.1 Pro · 89.2%

WebArena-VerifiedAgents48.8%#3

best: Qwen3.8-27B · 64.8%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Fine-tune it

Permissive

QLoRA19.0 GB1× RTX 3090 24GB

LoRA58.2 GB1× A100 80GB

Full fine-tune434.0 GB4× H200 141GB

QLoRA SFT on ~10k samples ≈ $13.91 (1× RTX 3090 24GB)

Plan a fine-tune →

Qwen3.6 family

Elo progression across releases

Qwen3.6-PlusApr 20262290.1 Qwen3.6-35B-A3BApr 20261973.2 Qwen3.6-27BApr 20262027.4

Resources

HFQwen/Qwen3.6-27B GHQwenLM/Qwen3.6 DOCShttps://huggingface.co/Qwen/Qwen3.6-27B APIAvailable via Alibaba API

API price $0.6/$3.6 · each benchmark row carries its own source badge (see methodology)