Alibaba launches 2.4T-parameter Qwen3.8-Max via API
Alibaba's new flagship is a 2.4-trillion-parameter mixture-of-experts model aimed at long-horizon agent work, with open weights promised next week.
A Max-class model, priced to undercut
Alibaba's Qwen team published Qwen3.8-Max on Monday, describing it as the most capable model the family has shipped. The company puts the total size at 2.4 trillion parameters in a mixture-of-experts configuration accepting text, image and video input, with a one-million-token context window — a maximum of 991,000 input tokens, dropping to 983,000 when reasoning is enabled, and up to 131,000 output tokens. Alibaba has not published an activated-parameter figure in its own materials; The Decoder reports roughly 95 billion active per query, which would leave inference cost far below the headline number.
International API pricing is $2.00 per million input tokens and $6.00 per million output tokens, with cached input reads at $0.25. Chinese pricing is set at ¥12 and ¥36 per million tokens, with cache hits at ¥1.5. The model is live on QwenCloud behind both OpenAI Chat Completions and Anthropic-compatible endpoints.
Long-horizon tasks as the pitch
Alibaba built its launch around multi-day autonomy rather than single-turn scores. Case studies include a command-line tool the model developed over 16 days across 265 commits and 127 pull requests, a roughly five-day research-paper reproduction spanning 33 GPU training jobs, and a multimodal competition entry scoring 0.853 that placed ahead of 458 of 526 human teams. On Terminal-Bench 2.1 the model scores 86.6, behind GPT-5.6 Sol at 88.8 and ahead of Anthropic's 84.6; Alibaba reports 93 on PaperBench.
Third-party boards moved quickly. LMArena placed Qwen3.8-Max fifth in Text Arena at 1,496 points and first in its medicine and healthcare slice, second in Vision Arena at 1,305 — thirteen points behind Claude Fable 5 (High) — and said the release reshaped the cost-performance frontier in Frontend Code Arena.
Why it matters
Qwen-Max has always been the closed tier of an otherwise open family, so the notable commitment here is Alibaba's statement that weights will ship next week on Hugging Face and ModelScope alongside a smaller Qwen3.8-27B checkpoint intended for on-premise hardware. Those weights are not public yet, and until they are this is an API launch. But if the release lands as described, a frontier-class model at roughly a third of Western flagship pricing — with a downloadable variant behind it — would put direct pressure on how closed labs price agentic workloads.