Alibaba's Qwen team pushed out Qwen3.8-Max, a mixture-of-experts model with roughly 2.4 trillion total parameters, of which only about 95 billion activate for any given query. That activation ratio matters more than the headline parameter count: it is the mechanism that lets a model this large stay usable and affordable to serve, rather than requiring a proportionally enormous amount of compute per request. The hosted API is live now through Alibaba's QwenCloud and is built to be a drop-in replacement for OpenAI- or DashScope-style integrations, meaning teams already wired up to one of those APIs can point their base URL and model ID at Qwen3.8-Max without rewriting client code. Alibaba has said the weights themselves will follow on Hugging Face and ModelScope within the week, which would make this the first model in the "Max" performance tier -- Alibaba's designation for its top-of-line, frontier-competitive models -- to ever ship as open weights rather than staying locked behind a hosted API only. On benchmarks Alibaba is citing, Qwen3.8-Max lands ahead of Anthropic's Claude Opus 4.8 and Claude Fable 5 on Terminal-Bench 2.1 (a test of how well a model handles real command-line and coding tasks), while trailing OpenAI's GPT-5.6 Sol at the very top end. It also leads on PaperBench and several vision-heavy benchmarks like OSWorld and document understanding. For developers, the practical implication is a second credible option (alongside Moonshot's recently released Kimi K3) for running a frontier-class coding and reasoning model on your own infrastructure once the weights land, without being locked into a single vendor's hosted pricing or rate limits.