Qwen3.6-27B
Qwen3.6-27B is a dense 27-billion-parameter multimodal model in the Qwen3.6 series, supporting both vision-language thinking and non-thinking modes in a single unified checkpoint. The 64-layer language model uses a hybrid layout of 16 repeats of (3 × Gated DeltaNet → FFN, 1 × Gated Attention → FFN) with hidden dim 5120 and FFN intermediate 17408 — Gated DeltaNet has 48/16 heads for V/QK (head dim 128) and Gated Attention has 24/4 heads for Q/KV (head dim 256). It supports a native 262,144-token context extensible to ~1,010,000 via YaRN and is trained with multi-token prediction. The release delivers flagship-level agentic coding, surpassing the previous-generation open-source flagship Qwen3.5-397B-A17B (397B total / 17B active) on every major coding benchmark including SWE-bench Verified (77.2), SWE-bench Pro (53.5), Terminal-Bench 2.0 (59.3), and SkillsBench (48.2), and reaches 87.8 on GPQA Diamond. Released as open weights under Apache 2.0; accessible via Qwen Studio with the Alibaba Cloud Model Studio API coming soon.
Benchmark results
| Benchmark | Score | Tags | Source |
|---|---|---|---|
| AIME 2026 | 94.1% | self-reported llm-stats | link → |
| AndroidWorld | 70.3% | self-reported llm-stats | link → |
| C-Eval | 91.4% | self-reported llm-stats | link → |
| CC-OCR | 81.2% | self-reported llm-stats | link → |
| CharXiv-R | 78.4% | self-reported llm-stats | link → |
| Claw-Eval | 60.6% | self-reported llm-stats | link → |
| CountBench | 97.8% | self-reported llm-stats | link → |
| DynaMath | 85.6% | self-reported llm-stats | link → |
| EmbSpatialBench | 84.6% | self-reported llm-stats | link → |
| ERQA | 62.5% | self-reported llm-stats | link → |
| GPQA | 87.8% | self-reported llm-stats | link → |
| HMMT 2025 | 93.8% | self-reported llm-stats | link → |
| HMMT Feb 26 | 84.3% | self-reported llm-stats | link → |
| HMMT25 | 90.7% | self-reported llm-stats | link → |
| Humanity's Last Exam | 24.0% | self-reported llm-stats | link → |
| IMO-AnswerBench | 80.8% | self-reported llm-stats | link → |
| LiveCodeBench v6 | 83.9% | self-reported llm-stats | link → |
| MathVista-Mini | 87.4% | self-reported llm-stats | link → |
| MLVU | 86.6% | self-reported llm-stats | link → |
| MMBench-V1.1 | 92.3% | self-reported llm-stats | link → |
| MMLU-Pro | 86.2% | self-reported llm-stats | link → |
| MMLU-Redux | 93.5% | self-reported llm-stats | link → |
| MMMU | 82.9% | self-reported llm-stats | link → |
| MMMU-Pro | 75.8% | self-reported llm-stats | link → |
| MMStar | 81.4% | self-reported llm-stats | link → |
| MVBench | 75.5% | self-reported llm-stats | link → |
| NL2Repo | 36.2% | self-reported llm-stats | link → |
| OCRBench | 89.4% | self-reported llm-stats | link → |
| QwenWebBench | 1,487 | self-reported llm-stats | link → |
| RealWorldQA | 84.1% | self-reported llm-stats | link → |
| RefCOCO-avg | 92.5% | self-reported llm-stats | link → |
| RefSpatialBench | 70.0% | self-reported llm-stats | link → |
| SimpleVQA | 56.1% | self-reported llm-stats | link → |
| SkillsBench | 48.2% | self-reported llm-stats | link → |
| SuperGPQA | 66.0% | self-reported llm-stats | link → |
| SWE-bench Multilingual | 71.3% | self-reported llm-stats | link → |
| SWE-Bench Pro | 53.5% | self-reported llm-stats | link → |
| SWE-Bench Verified | 77.2% | self-reported llm-stats | link → |
| Terminal-Bench 2.0 | 59.3% | self-reported llm-stats | link → |
| V* | 94.7% | self-reported llm-stats | link → |
| VideoMME w sub. | 87.7% | self-reported llm-stats | link → |
| VideoMMMU | 84.4% | self-reported llm-stats | link → |
| VLMsAreBlind | 97.0% | self-reported llm-stats | link → |
| ZClawBench | 53.4% | self-reported llm-stats | link → |