Alibaba prices Qwen3.8-Max at $2 per 1M input tokens and $6 per 1M output tokens via its API, below Kimi K3's $3/1M input tokens and $15/1M output tokens (Henry Siu/The Information)

Alibaba Group Holding Ltd. unveiled its latest open-source large language model, Qwen3.8-Max, boasting an impressive 2.4 trillion parameters, seven times greater than its February release, Qwen3.5. This new model can process up to 1 million tokens in prompts, enabling it to analyze more than 200 pages of text or 100 hours of footage per request.

Developed by leveraging Nvidia's Gated DeltaNet attention mechanism, this model can respond with up to 131,000 tokens and scales linearly as prompt size increases, reducing hardware usage. Alibaba tested Qwen3.8-Max across various tasks, including a 16-day coding project and a chip design optimization requiring over 500 steps, scoring 1,668 points on a benchmark measuring LLM skills in interface development.

This performance puts it just 37 points behind Anthropic PBC's latest model, Claude Opus 5. Qwen3.8-Max is currently available on Alibaba's cloud platform and will be open-sourced next week, alongside a more hardware-efficient version called Qwen3.8-27B.