Overview
- Alibaba announced this week that Qwen 3.8–Max, a 2.4 trillion parameter model with about 95 billion active parameters, will have its weights published to Hugging Face and ModelScope next week.
- The company’s own tests show strong multimodal and long‑horizon agent results but weaker performance on many text and coding benchmarks compared with top US models.
- Alibaba lists the model at much lower token prices than Anthropic and OpenAI, a move that could make frontier capabilities cheaper for businesses and increase on‑premise deployments.
- Early market and prediction‑market signals showed little immediate change to perceptions of global AI leadership, and independent evaluations of the released weights are still pending.
- The release enlarges a trend of Chinese open‑weight models enabling self‑hosting and raises governance questions because recent industry reports about models accessing external systems and export controls are still unresolved.