According to the official announcement, Alibaba Qwen has released an upgraded version of its flagship model, Qwen3.8-Max-0902, with parameters remaining at 2.4T and context length capped at 1M tokens. The model underwent post-training focusing on programming, complex workflow execution, and long-horizon Agent tasks. According to benchmark data released by Alibaba, Qwen3.8-Max-0902 outperforms the original version across all eight programming evaluation metrics: TerminalBench 3.0 increased from 11.3 to 29.0, DeepSWE 1.1 rose from 56.6 to 69.3, QwenSWEbench V2 improved from 55.1 to 70.0, and JobBench climbed from 53.4 to 64.0.

Compared to Claude Opus 5, Qwen3.8-Max-0902 still lags in most programming benchmarks but surpasses it in three key evaluations: MLS-Bench-Lite, SWE-Atlas QnA, and QwenSWEbench V2. Multimodal performance saw only marginal gains of 0.4 to 3 points across four benchmarks, indicating that this iteration primarily enhances Agent-level code execution and task orchestration capabilities. All results are internal benchmarks conducted by Alibaba. Qwen3.8-Max-0902 is now available via QwenCloud API with unchanged pricing: $2 per million input tokens and $6 per million output tokens.