Key points
- LongCat-2.5 weights are public under the MIT license, which permits commercial use, modification and redistribution, one week after the model's official unveiling.
- Mixture-of-experts design: 1.6 trillion total parameters, about 48 billion active per token; API handles up to 1 million tokens of context.
- Meituan reports 88.4 on Terminal Bench v2.1 versus 89.1 for Claude Opus 5, but 38.5 on the harder v4.0 versus Opus 5's 53.9.
- API accepts OpenAI and Anthropic formats and works with Claude Code and Codex; launch pricing is roughly €0.27 per million input tokens.
Why it matters for marketers: Teams building agent workflows get a permissively licensed, low-priced model that plugs into existing coding harnesses, though benchmark gaps and self-reported results mean it needs testing on real tasks.
Meituan's LongCat team has published the weights for LongCat-2.5, a model designed for agents that work through long, multi-step tasks, with coding as a focus. According to Meituan, it was trained entirely on Chinese hardware and comes close to leading Western models on several benchmarks. It accepts text, images and video and replies in text, with reasoning that can be switched on or off.
Results are mixed. Meituan reports 88.4 on Terminal Bench v2.1, just behind Claude Opus 5 at 89.1. On the harder v4.0 variant it scores 38.5, slightly above GPT-5.6 Sol but well under Opus 5's 53.9. It posts 91.9 on BrowseComp, the top score in Meituan's table, though the company ran that test in its own environment. On DeepSWE it reaches 64.6, behind most models listed.
For marketing and content teams, the practical points are licensing and access. The MIT license allows commercial use and modification, and the API supports both OpenAI and Anthropic formats, so it can slot into agent harnesses such as Claude Code and Codex. A limited-time promotional price puts input at about €0.27 per million tokens and output at about €1.07.
Meituan's self-hosting guide describes a cluster of 194 Huawei Ascend chips and offers no guidance for Nvidia GPUs. Teams considering it should run their own tests on real workflows, since the benchmark numbers come from Meituan, and check whether the promotional pricing holds once the launch period ends.
Sources
This brief is our own summary. Read the full reporting at the source.