Perceive Beating AI News Flash: Alibaba's One Thousand Questions has just upgraded its flagship Qwen3.8-Max, which was released just a month ago, to introduce Qwen3.8-Max-0902. The parameter scale remains the same, with 2.4T parameters and a context of 1 million tokens. This time, the focus is on post-training for Coding and Cowork, specifically enhancing programming, complex workflows, and long-duration Agent tasks.
In the comparison table released by Alibaba, the 0902 model outperforms the original version in all 8 programming benchmarks. TerminalBench 3.0 increased from 11.3 to 29.0, DeepSWE 1.1 from 56.6 to 69.3, and QwenSWEbench V2 from 55.1 to 70.0. The professional task JobBench also rose from 53.4 to 64.0.
Compared to Claude Opus 5, the 0902 model still loses in most programming tasks, but it surpasses the opponent in MLS-Bench-Lite, SWE-Atlas QnA, and QwenSWEbench V2. The scores in the four multimodal tasks only increased by 0.4 to 3 points compared to the original version, indicating that the focus of this upgrade is clearly to make the Agent more proficient in coding and tasks. All the mentioned achievements are currently from Alibaba's internal testing.
Qwen3.8-Max-0902 has now been launched on the QwenCloud API, maintaining the same price of $2 per input and $6 per output per million tokens.
