THE CRUNCH
Meituan's LongCat team has published the weights of LongCat-2.5, a 1.6-trillion-parameter model designed for AI agents that carry out lengthy, multi-step work on their own, with coding the main focus. The release comes one week after the model's official unveiling. According to Meituan, it was trained entirely on Chinese hardware and performs near the level of leading Western models on several benchmarks.
LongCat-2.5 uses a mixture-of-experts design, so only part of the network runs for each token: roughly 48 billion of the 1.6 trillion parameters are active at a time. Via the API it can take in up to one million tokens in a single request and produce replies of up to 128,000 tokens. It reads text, images and video, answers in text, and offers a switch for turning reasoning on or off. Rather than attaching a separate image encoder, it turns images into tokens that sit alongside text in the same processing stream.
To train the model with reinforcement learning, the team drew on genuine software tasks taken from GitHub pull requests plus reconstructions of agent mistakes. They also took precautions to confirm the model was genuinely solving the tasks: training environments had internet access blocked, Git history and cached files removed, and a language model checked solutions for tricks.
The benchmark picture is mixed. The LongCat team reports 88.4 points on Terminal Bench v2.1, which measures how well agents solve command-line tasks, with Meituan listing Claude Opus 5 at 89.1 and Deepseek-V4.1-Flash leading at 90.6. On the harder Terminal Bench v4.0 variant it scores 38.5, slightly above GPT-5.6 Sol's 37.3 but well below Opus 5's 53.9. On DeepSWE, which tests bug-fixing in real code repositories, it manages 64.6, behind every other listed model except Claude Opus 4.8. Its 91.9 on the web search benchmark BrowseComp is the highest in the table, though Meituan ran that test in its own environment.
The weights ship under the MIT licence, which permits commercial use, modification and redistribution, and the model can be tried at longcat.ai. The API speaks both the OpenAI and Anthropic formats and works with agent harnesses such as Claude Code, Codex and Hermes. At launch, Meituan is charging a limited-time promotional price equivalent to roughly €0.27 per million input tokens and €1.07 per million output tokens.


