Qwen's Max flagship goes open-weights: 2.4T MoE now on Hugging Face
Alibaba's first downloadable Max-class model: 95B active of 2.4T, ~1M context — but the license isn't Apache. Plus new agent tooling in the AI SDK.
Copy markdown
The first Max-class Qwen you can run
Qwen3.8-2.4T-A95B is a mixture-of-experts model that activates 95B of 2.4T params per token (512 experts, 10 routed + 1 shared) — the first time Alibaba has open-weighted a Max-tier model. Weights are live on Hugging Face and ModelScope in Transformers format.
Read the license before you ship
It's open weights, not open source: the drop carries a custom “Qwen3.8-Max” license rather than the Apache-2.0 that covers the smaller Qwen3.8-27B. Check the commercial and redistribution terms before you build a product on it.
Frontier-class scores, with a reasoning dial
Reported numbers land high: 92.6 GPQA Diamond, 93.0 PaperBench, 86.6 Terminal-Bench 2.1, and a weaker 67.7 on SWE-bench Pro. A reasoning_effort control (low/medium/xhigh) trades latency for depth on each call.
~1M context, or rent it by the token
Context runs 262K native and stretches toward ~1M for whole-repo and long-horizon agent work. Don't want to host 2.4T of weights? The qwen3.8-max API is OpenAI- and Anthropic-compatible at $2/$6 per million tokens, $0.25 cached.
Elsewhere: any agent harness, now in the AI SDK
Vercel shipped @ai-sdk/harness-acp and @ai-sdk/harness-grok-build, so you can drive any Agent Client Protocol harness — or Grok Build — through the AI SDK's HarnessAgent interface. One loop, swappable agent backends.