Qwen
Alibaba's Qwen model family โ open-weight leaders (Qwen3.6-27B) and closed-weight pivot (Qwen3.7 Max/Plus); SEA-LION regional variant
Qwen
Entity page for Qwen โ Alibaba's model family and one of the most-covered open-weight lines in this knowledge base.
Overview
Qwen (้ไนๅ้ฎ) is developed by Alibaba's Qwen team. The 2026 story has two chapters:
- Open-weight leadership โ Qwen3.6-27B (Apache 2.0) as frontier-adjacent dense model
- Closed-weight pivot โ Qwen3.7 Max/Plus competing with Opus and GPT-5.5 via API
Current models (Sep 2026)
| Model | Type | Highlights |
|---|---|---|
| Qwen3.8-Max (2.4T-A95B) | Open MoE (FP8) | 2.4T total, 95B active, 86.6% Terminal Bench 2.1, 56.6% DeepSWE 1.1, 93.0% PaperBench, $2/$6 API, 1M context, first open-weight Max-class |
| Qwen3.8-Flash-Next | Open MoE (Experimental) | Qwen4 architecture preview: QSA, Gated Residual, N-gram Embedding, 6B active + 51B n-gram, 58.7% DeepSWE |
| Qwen3.8-27B | Open dense | 27B, hybrid Gated DeltaNet + Gated Attention, native vision-language, 262K context, Apache 2.0, 61.7 SWE-bench Pro, 84.3 OSWorld-Verified |
| Qwen3.7 Max | Closed API | 90/100 BenchLM, 69.7 Terminal-Bench 2.0, 44.5 Apex, 80.4% SWE-Verified, 1M context |
| Qwen-AgentWorld-397B-A17B | Open MoE | 58.71 AgentWorldBench (beats GPT-5.4), 7-domain LWM, 256K context |
| Qwen-AgentWorld-35B-A3B | Open MoE | 56.39 AgentWorldBench, 7-domain LWM, Apache 2.0 |
| Qwen3.7 Plus | Closed multimodal | Vision+video agent, $0.32/$1.28 |
| Qwen3.6-27B | Open dense | 77.2% SWE-Bench Verified; beats 397B MoE sibling |
| Qwen3.6-35B-A3B | Open MoE | 35B total, 3B active; thinking preservation |
| Qwen-SEA-LION-v4.5-27B | Regional (AISG) | SEA languages, distilled from Qwen3.5-397B |
Strategic shifts
Open-weight Max-class arrives (AugโSep 2026)
Qwen3.8-Max (2.4T-A95B) is the first trillion-parameter MoE released with open weights (FP8). This reverses the closed-weight pivot of Qwen3.7 and democratizes frontier-level agentic coding. At $2/$6 per million tokens, it's 5ร cheaper on input and 8.3ร cheaper on output than US frontier models. The accompanying Qwen3.8-Flash-Next previews the Qwen4 architecture (QSA, Gated Residual, N-gram Embedding).
โ Qwen3 8 Max First Open Source Flagship Hybrid Deltanet Moe Coding Cowork 2026 09 18
Dense beats MoE (Apr 2026)
Qwen3.6-27B proved a 27B dense model can outperform Alibaba's own 397B MoE on agentic coding โ challenging the "bigger MoE always wins" narrative.
โ Qwen36 27b Dense Beats Moe Agentic Coding Analysis 2026 06 03
Open โ closed pivot (MayโJun 2026)
Qwen3.7 Max moves to proprietary API-only distribution, competing for enterprise revenue against Claude and GPT. Open-weight community goodwill from Qwen3.6 does not automatically transfer.
โ Qwen37 Max Plus Closed Weight Frontier Agent Era 2026 06 16
Caveat: Qwen3.7 Max verbosity (~4ร median output tokens) narrows stated cost advantages.
Key benchmarks
| Model | SWE-bench Pro | SWE-bench Verified | Terminal-Bench 2.1 | OSWorld-Verified | Notes |
|---|---|---|---|---|---|
| Qwen3.8-Omni-Flash | โ | โ | โ | โ | Native omnimodal (text+image+audio+video), 1M context, $0.15/$0.47, WildClawBench-MM 71.0 (+106%), LongAudioSpan 82.7, Qwen-MM-Plugins |
| Qwen3.8-Max | 67.7% | โ | 86.6 | โ | 2.4T MoE, 95B active, open weights (FP8), 93.0% PaperBench, 56.6% DeepSWE |
| Qwen3.8-27B | 61.7% | โ | 73.0 | 84.3 | 27B dense, hybrid attention, native vision-language, Apache 2.0 |
| Qwen3.7 Max | 60.6% | 80.4% | 69.7 | โ | 35-hour autonomous kernel demo, 1M context |
| Qwen-AgentWorld-397B | โ | โ | โ | โ | 58.71 AgentWorldBench, 7-domain LWM |
| Qwen3.6-27B | โ | 77.2% | โ | โ | Perfect 100/100 tool-eval-bench |
| Qwen3.6-35B-A3B | โ | competitive | โ | โ | Thinking preservation |
Key articles
Related
- Concepts: Frontier Models, Mixture Of Experts, Agentic Coding
- Compared to: DeepSeek-V4-Pro, Gemma 4, GLM-5.2, Claude Opus
- Deployment: Howto Multi Model Routing Layer, Howto Vllm Deployment Guide
Link map
Solid arrows: links from this page. Dashed arrows: pages that link here.
๐ Referenced by
- ๐ Journal Entry - October 6, 20262026-10-06T00:00:00.000Z
- ๐ Journal Entry - October 5, 20262026-10-05T00:00:00.000Z
- ๐ September 18: Qwen3.8-Max โ First Open-Source Flagship, Hybrid DeltaNet MoE, and the 10-Day Autonomous Coding Benchmark2026-09-18T00:00:00.000Z
- ๐ August 5: Qwen3.8-Max โ 2.4T Parameters, 16 Days of Autonomous Coding, and the Open-Weight Frontier Expands2026-08-05T00:00:00.000Z
- ๐ July 1: Qwen3.7-Max and the Language World Model Era2026-07-01T00:00:00.000Z
- ๐Wiki Index2026-06-17T00:00:00.000Z
- ๐Wiki Log2026-06-17T00:00:00.000Z