Meituan, the Chinese company best known for delivering food across its home market, has released a new AI model built for a very different kind of long job: staying with a complex task for hours, not just answering a single prompt. Called LongCat-2.5, it was released on November 2, 2026, after a September API preview. It is described as a native multimodal model for long-horizon agentic tasks — multi-step jobs in which the AI acts on its own, such as editing a complex codebase or building a complete project. That focus is what makes this launch different from a routine model update.
The details
LongCat-2.5 uses a mixture-of-experts design with 1.6 trillion total parameters, of which about 48 billion are
active per token. That scale is paired with a 1M-token context window and up to 128K tokens of output, according to API documentation reports — room for very long inputs and very long working sessions. In practice, that means long
coding work, rather than short answers. It is designed to work with agent tools including Claude Code, Codex and Hermes, and it is live through Meituan's LongCat chat service and API, with GitHub and Hugging Face releases planned as “coming soon”. It follows LongCat-2.0, which debuted in June 2026.
Why it matters
The most important part of this story may not be the parameter count. Meituan says it optimized long-context training and W4A4 inference — 4-bit weights and activations — for domestic Chinese AI accelerators, each with under 80 GB of memory. It has not named the chip vendor. That is the company's own claim, and it sits against a backdrop of US export restrictions on advanced AI chips. If the claim holds, it suggests a Chinese company can train and run a very large model without relying on US hardware. For
AI agents that must work through long, multi-step jobs, that combination of model and domestic hardware is the point: the chip race and the model race are now the same race.
A short take
Read the benchmarks with care. They are Meituan's own, not independently verified. The company's table reports 88.4 on Terminal-Bench v2.1, 51.8 on AutomationBench and 91.9 on BrowseComp, and claims competitiveness with Claude Opus variants on some tests. But on Terminal-Bench v4.0, its own table shows 38.5 versus 53.9 for Claude Opus 5. The honest reading is that LongCat-2.5 looks built for long, practical work, but the strongest claims still need independent testing. If it delivers on that promise, it will matter far beyond one company's product line.
Comments
Post a Comment