Tier 3 Practical open-weight
What the publisher claims: open-weight, open-source
The one-liner: Broadly deployable open-weight family whose top tier now carries a custom license.
The six questions
Weights are downloadable across the whole family, from small dense releases up to the 2.4T-A95B Max-tier MoE, in multiple precisions.
Most of the family is clean Apache 2.0, but the flagship open-weight Max tier ships under the custom Qwen3.8-Max License with model-name display and separate-license conditions, and older Qwen2.5 checkpoints retain a custom license, so permission has to be checked at the exact checkpoint.
Highly portable: dense checkpoints run on a single consumer GPU while the 397B and 2.4T MoE releases run on standard multi-node NVIDIA infrastructure with vLLM and SGLang.
Model reports describe architecture, parameter counts, token counts, context windows, and post-training stages, and reporting on the Qwen3.8 release records training data, methods, and knowledge cutoff as undisclosed; no training corpus or reproducible recipe is published.
Limited to moderate: the published reports describe what was built, not how to rebuild it.
Alibaba's Qwen releases land on Hugging Face and GitHub where outside tooling pull requests are sometimes merged and community feedback is read, but the model and training recipe are developed internally by the Qwen team.
License
Apache 2.0 across Qwen3, Qwen3.5, and Qwen3.8-27B; custom Qwen3.8-Max License for the 2.4T-A95B Max tier; older Qwen2.5 checkpoints retain a custom Qwen license — license source ↗
- Qwen3.8-Max License requires displaying the model name above 100M MAU or $20M monthly revenue
- Qwen3.8-Max License requires a separate Qwen license for Model-as-a-Service or AI Work Assistant businesses above $50M TTM aggregate revenue
- Qwen2.5 checkpoints carry an MAU threshold and derived-model attribution conditions
What is actually released
- weights (Qwen3, Qwen3.5 up to 397B-A17B, Qwen3.8-27B, Qwen3.8-2.4T-A95B)
- quantized and full-precision builds
- model reports with architecture, parameter counts, and context windows
- training-stage descriptions and evaluation results
Why it matters
Qwen shows how a model can be highly portable and permissively licensed at the small end while the frontier tier moves to revenue-conditioned terms, which is the clearest 2026 case of a family split across two license regimes.
Sources
- https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/blob/main/LICENSE
- https://huggingface.co/Qwen/Qwen3.5-397B-A17B/blob/main/LICENSE
- https://github.com/QwenLM/Qwen3.8
- https://www.deeplearning.ai/the-batch/qwen3-8-max-lands-with-a-bang
- https://llm-stats.com/blog/research/qwen3-8-max-open-weights
Derived from the openness dataset behind But Is It Open? — snapshot 2026-11-07 ↗. Scores are qualitative readings of the cited evidence, not verdicts; tier measures openness, not intelligence.