Alibaba has revised its largest language model: the „0902“ update, released on 2 September, delivers better results across all eight coding tests the company ran – with unchanged architecture and pricing.
Only the post-training changed
Technically, little is new. The model still runs on roughly 2.4 trillion parameters in a mixture-of-experts design with a one-million-token context window. Pricing is unchanged too – two US dollars per million input tokens and six per million output tokens. According to the Qwen team, only the post-training was improved, meaning the fine-tuning that follows the base training run.
Gains focused on programming
The clearest jump comes in coding and office tasks. On the independent Code Arena WebDev leaderboard, which scores web-development work, the 0902 version ranks first with 1,691 points – just ahead of Anthropic’s Claude Opus 5 Max and the previous release. Several coding benchmarks more than doubled their scores, the company says. Those figures come mainly from Alibaba’s own measurements and should be read with that in mind.
- Rank 1 on Code Arena WebDev (1,691 points)
- All eight coding tests improved
- Price and context window unchanged
Claude still leads on several agent and reasoning benchmarks. For developers in Germany, Qwen3.8-Max remains chiefly a price-competitive alternative.
Sources: OpenRouter · DataCamp · Wikipedia
