Alibaba's Qwen3.8-Max nearly matches Anthropic on coding and visual tests — and the weights arrive next week

Alibaba released Qwen3.8-Max, its largest generative AI model to date: a system trained on text and images that can write code, answer questions, and analyze pictures, comparable to ChatGPT or Claude. The 2.4-trillion-parameter model — where each parameter is one trained setting the system learned during training — ranks just behind Anthropic's top system and a small group of other top models on the Arena.AI crowdsourced text leaderboard, a public ranking where users compare anonymous model outputs head-to-head.

On Alibaba's own tests, the model matches or beats Anthropic's flagship on frontend coding and visual analysis. For coding, only two Claude Opus models and Moonshot's Kimi K3 reportedly score higher; for visual analysis, only one system outperforms it.

Alibaba said it will release the model's trained settings, called weights, next week under an open-weight license, meaning developers can run, study, and fine-tune the system on their own hardware. The release returns Alibaba to open-weight distribution after a brief pivot toward closed products, and it puts a near-frontier Chinese model into the global developer stack that US export controls cannot easily reach once the weights are in the wild.

Caveats remain. Alibaba's coding and image-test results are self-reported, with no independent review yet, and parameter counts for OpenAI and Anthropic's top systems are not disclosed, so the US-China comparison is asymmetric. The leaderboard peg is firmer ground, but the open-weight release next week is the move that will outlast any single benchmark cycle.