Qwen3.8-Max-0902

Qwen3.8-Max-0902, released September 2, 2026, was a post-trained refresh of the Qwen3.8-Max Alibaba had shipped a month earlier rather than a new model underneath it. The 2.4-trillion-parameter base and the 1M-token context window were unchanged; what the lab changed was the post-training, which it said had been extended on coding and on Cowork, the agentic office work its own CoWorkBench measures. The dated suffix marked the checkpoint the way DeepSeek had been naming its own refreshes since V3-0324, and it kept the older checkpoint addressable instead of replacing it in place.

The launch table was set against the August checkpoint rather than against rival labs, and the coding rows carried most of the movement. Qwen3.8-Max-0902 scored 29.0% on Terminal-Bench 3.0 where the older checkpoint had managed 11.3%, 69.3% on DeepSWE 1.1 against 56.6%, and 64.9% on NL2Repo-Bench against 55.9%; on QwenSWEBench V2, Alibaba's own harder in-house set, it posted 70.0% to the earlier model's 55.1%, ahead of the 68.0% and 67.1% it reported for Claude Opus 5 and Fable 5. Agent work moved less. CoWorkBench went from 74.8% to 76.1% and Toolathlon-Verified from 72.5% to 73.3%, both still behind Claude Opus 5, while JobBench jumped from 53.4% to 64.0%. Multimodal scores were flat, at 82.7% on MMMU-Pro against 82.3%.

Pricing was $2 per million input tokens and $6 per million output, with cache hits at $0.17 explicit and $0.25 implicit — the first time the Max tier had been quoted on QwenCloud, the API Alibaba had introduced with Qwen3.8-Flash the previous week, rather than through Alibaba Cloud Model Studio. It launched API-only, with no weights published for the refreshed checkpoint, leaving the August base model the most recent Qwen3.8 weights available.