Qwen3.8-Max ships: 2.4T MoE, 1M context, $2/$6 per Mtok
Alibaba's new flagship roughly doubles its agentic scores, ships OpenAI-compatible and on Vercel today; open weights and a self-hostable 27B land next week.
Copy markdown
Call it today, OpenAI-compatible
qwen3.8-max is live now on Alibaba's DashScope (OpenAI- and DashScope-compatible endpoints) and on Vercel's AI Gateway as alibaba/qwen3.8-max. It runs $2/$6 per Mtok in/out with cached input as low as $0.17, a 1M-token window (991K in, 131K out) and 2M tok/min limits — a cheap drop-in for anything already speaking the OpenAI API.
The agentic jump is the real story
Against Qwen3.7 the agentic scores roughly doubled: DeepSWE 1.1 went 21.6 to 56.6, FrontierSWE 40.7 to 73.5, JobBench 31.3 to 53.4. On Terminal-Bench 2.1 it posts 86.6, edging Claude Opus 4.8 (84.6) and trailing GPT-5.6 Sol (88.8). It's a far stronger tool-caller than the last Qwen.
Where it still trails — know before you route
It is not the top pure coder: SWE-bench Pro lands at 67.7 against Fable 5's 80.0, and FrontierSWE 73.5 vs 88.8. It's strong on agentic/tool-use, computer control and docs (OSWorld-Verified 86.1, OmniDocBench 92.1), but for hard patch-writing you may still route elsewhere.
Open weights next week, plus a 27B you can self-host
Alibaba says the weights land next week on Hugging Face and ModelScope — its first open Max-class release. The 2.4T flagship needs multi-node datacenter hardware, but a companion Qwen3.8-27B goes open too and fits standard GPUs. Weekend build: fine-tune the 27B on your own agent traces.
Batteries-included API
The hosted model ships with function calling, structured outputs, batches, prefix completion and fine-tuning, plus built-in code_interpreter, web_search, web_extractor and image-search tools. Inputs cover text, image and video (text out), so you can wire multimodal agents without bolting on separate tool services.