Meituan releases LongCat-2.5, a 1.6-trillion-parameter agent model

LongCat-2.5 builds on its predecessor LongCat-2.0, which Meituan says it trained in late June entirely on a cluster of more than 50,000 Chinese AI chips. Before that release, LongCat-2.0 ran anonymously as "Owl Alpha" on OpenRouter. The new model adds image understanding by converting images into tokens processed alongside text, drawing on the research model LongCat-Next, rather than bolting on a separate image encoder.

For reinforcement learning, the team used real software tasks from GitHub pull requests and recreated agent failures. To ensure the model actually solved tasks during training, developers blocked internet access, removed Git history and cached files, and had a language model check solutions for shortcuts. They also distilled capabilities from specialized teacher models for agent work, reasoning, and dialog.

The benchmark picture is mixed. On Terminal Bench v2.1, LongCat-2.5's 88.4 sits behind Claude Opus 5 (89.1), GPT-5.6 Sol (88.8), and Deepseek-V4.1-Flash (90.6). On DeepSWE, it jumps from 9.7 to 64.6 points, placing it behind every other model in the table except Claude Opus 4.8. On the harder Terminal Bench v4.0, it scores 38.5, above GPT-5.6 Sol's 37.3 but well below Opus 5's 53.9. It posts 91.9 on BrowseComp, the highest in the table, though Meituan ran that test in its own environment. The release follows other Chinese open-weight models like Kimi K3.1 and GLM-5.3 that have largely caught up with top US models on common benchmarks.