Chinese delivery giant Meituan has put LongCat-2.5 online for download, a huge model aimed at AI agents that grind through long coding jobs. Anyone can download it, change it and use it commercially, because it carries the MIT licence.
Only about 48 billion parameters are active for each token, since the model is a mixture-of-experts design. It reads images as well as text, and its image understanding is new compared with LongCat-2.0, the version before it.
On Meituan’s own model card, LongCat-2.5 scores 88.4 on Terminal Bench v2.1, against 89.1 for Claude Opus 5 and 88.8 for GPT-5.6 Sol. The harder Terminal Bench v4.0 shows a wider gap: 38.5 for LongCat, 53.9 for Opus 5. On DeepSWE, which tests fixes in real code repositories, it rose from 9.7 for LongCat-2.0 to 64.6, still below GPT-5.6 Sol’s 72.7. These are Meituan’s figures, and the company marks its own tests as indicative only.
Meituan trained it with reinforcement learning on genuine engineering work, and says it built in checks so the model cannot score by gaming the reward.
A second release shrinks the expert layers to 4-bit numbers, a format that Chinese-made AI accelerators handle natively. Against an 8-bit setup, its stress tests show 43.3% higher prompt-processing throughput and 42.77% higher generation throughput, with accuracy it describes as comparable to the full model. The card points NPU users to the SGLang-FluentLLM project for deployment.
For developers, the licence is the practical part: a model near the top of agent benchmarks that can be self-hosted without a usage agreement. The size is the catch: a model this large needs a serious GPU or accelerator cluster to run.
Meituan also runs a chat site for LongCat.
