Qwen 3.6

weights

open(Apache-2.0 for open variants e.g. 27B/35B-A3B)

data

closed

code

partial(inference/usage code on GitHub

license

Apache-2.0(most restrictive distributed SKU

Apache-2.0 across the distributed base variants, data closed and only inference code published, so 3. The load-bearing exclusion: Qwen3.7-Max is proprietary and served only through Alibaba Model Studio, so it is not part of the distributed set and does not cap the family. Including it would publish Qwen as 1/closed, which is what the rubric's multi_sku_rule exists to prevent - API-only tiers are a different product surface from the weights you can download. Qwen 2.5 scored 2 because its distributed 3B and 72B SKUs carried the Qwen License Agreement; from Qwen3 onward the distributed set is Apache. Replaces an earlier note that was a copy of the components string and recorded the license as `mixed`, a value the rubric forbids because it states that SKUs differ without recording the outcome.

• https://huggingface.co/api/models/Qwen/Qwen3.6-27B recorded 2026-07-29

  license apache-2.0, ungated, 55.6GB of weight files, 6,232,995 downloads last month

• https://github.com/QwenLM/Qwen3.6 recorded 2026-06-04

  flagship phase-C verification source

• https://www.marktechpost.com/2026/04/22/alibaba-qwen-team-releases-qwen3-6-27b-a-dense-open-weight-model-outperforming-397b-moe-on-agentic-coding-benchmarks/ recorded 2026-06-04

  flagship phase-C verification source

• https://llm-stats.com/models/qwen3.6-plus recorded 2026-06-04

  flagship phase-C verification source

• https://aimlapi.com/blog/qwen-3-6-series-alibabas-open-source-llm-revolution-in-2026 recorded 2026-06-04

  flagship phase-C verification source

Qwen family ~942M cumulative HF downloads (>50% of all global open source downloads per ATOM); Qwen2.5 was the dominant fine-tune base through 2025. The 72B base alone shows 68,655 downloads/mo (confirmed). Family reach unambiguously >10M; calibrated to Qwen 3.6 anchor (A5).

• https://huggingface.co/Qwen/Qwen2.5-72B recorded 2026-06-04

  live 72B base repo, 68,655 downloads/mo

Dense Qwen3.6-27B: SWE-bench Verified 77.2, Terminal-Bench 2.0 59.3 (reported matching Claude 4.5 Opus class), outperforming a 397B MoE on agentic coding. Flagship Qwen3.6-Plus reported matching Claude Opus 4.5 on SWE-bench/Terminal-Bench. Frontier-tier open family.

• https://www.marktechpost.com/2026/04/22/alibaba-qwen-team-releases-qwen3-6-27b-a-dense-open-weight-model-outperforming-397b-moe-on-agentic-coding-benchmarks/ recorded 2026-06-04

  flagship phase-C verification source

• https://www.caixinglobal.com/2026-04-02/alibaba-releases-qwen-36-plus-ai-model-with-enhanced-coding-capabilities-102430395.html recorded 2026-06-04

  flagship phase-C verification source