Update: Qwen3.8 open weights due this week — verify before you pull
Alibaba's 2.4T Qwen3.8-Max (1M context, $2/$6 per M) is live on API now; the open-weight 27B and Max hit Hugging Face this week, license undisclosed.
Copy markdown
The window's open, the repo isn't
Alibaba committed to shipping open weights for Qwen3.8-Max and a new Qwen3.8-27B the week of Aug 10, but no matching repo has appeared on Qwen's official Hugging Face org yet. Watch the org page directly, not mirrors or community reuploads.
What you can build on today: Qwen3.8-Max via API
The flagship is a 2.4T-parameter sparse MoE (~95B active) with a 1M-token context, live on QwenCloud at $2 per M input, $6 per M output, and $0.25 per M cached reads — frontier-scale context well under Opus/GPT-class pricing.
The coding numbers
Qwen3.8-Max posts PaperBench 93.0, Terminal Bench 2.1 at 86.6, and roughly 87% on SWE-bench, and ranks #4 in Frontend Code Arena — putting it in the conversation with Claude Opus 4.8 and GPT-5.6 Sol for agentic coding.
Read the license before you fine-tune
Terms are still undisclosed, and early readers flagged geographic restrictions (US, EU, UK, Korea) later softened to an 'authorization' process. Qwen3.6-27B shipped Apache-2.0, but that's precedent, not a promise for 3.8.
Elsewhere: Cursor's Auto Balance undercuts Opus 4.8
Cursor's updated Auto routing now claims to beat Claude Opus 4.8 on user satisfaction at 41% lower cost, and reach above-Fable satisfaction at 68% less — a default-on change to what your agent spends per task.