Meituan releases LongCat-2.5, a 1.6-trillion-parameter multimodal model built for agent work
Meituan's LongCat team introduced LongCat-2.5, a native multimodal mixture-of-experts model with 1.6 trillion total parameters aimed at long-running agent tasks such as coding whole projects, visual web building and 3D work, available in its chat app and through an API.
What happened
- LongCat's blog post, dated November 1, 2026, introduces LongCat-2.5 with about 1.6 trillion total parameters and roughly 48 billion active per token, building on LongCat-2.0.
- The team says it trained the model with large-scale reinforcement learning on long, multi-step tasks, and turns real agent failures into repeatable training tasks to close gaps.
- Images are turned into discrete tokens so the model handles pictures and text in one format, which the team says supports visual coding, design and 3D tasks.
- Meituan's newsroom says training ran entirely on Chinese-made AI chips, and that people can try the model at longcat.ai while developers can call it through an API launched at the same time.
- Meituan's own benchmark table places it between Claude Opus 4.8 and Claude Opus 5 on many tests; TechNode reports GitHub and Hugging Face releases are planned to follow.
Why it matters: It adds another large Chinese model aimed squarely at coding agents and office-style agent tasks, trained entirely on Chinese-made chips. The benchmark comparisons are Meituan's own, so wait for independent tests before switching tools.
Sources
- LongCat blog: Introducing LongCat-2.5 longcat.ai
- Meituan newsroom meituan.com
- TechNode technode.com
Checked November 2, 2026 against the sources above. Claims about performance or benchmarks are the company's own unless stated otherwise.