Meituan announces LongCat-2.5 with visual and agentic coding

Meituan describes LongCat-2.5 as an iteration on LongCat-2.0, with stronger native visual understanding and agentic coding.[2] The company says the model processes visual and textual inputs in a unified token-based format, and combines visual understanding with agentic coding for visual coding, creative design and 3D tasks.[2] Its stated focus extends to complex software work: Meituan says large-scale, long-horizon reinforcement learning is used to strengthen end-to-end engineering capabilities for complex codebases and full-project builds.[2] It also says real-world agent failure trajectories are converted into reproducible, verifiable training tasks.[2]

Meituan says LongCat-2.5 underwent long-context training with sequences of up to 1 million tokens and a total training volume of trillions of tokens.[1] The company also says the model was trained entirely on domestic computing hardware.[1] In Meituan’s benchmark table, LongCat-2.5 scored 88.4 on Terminal Bench v2.1, compared with 70.8 for LongCat-2.0, and 51.8 on AutomationBench, compared with 15.8 for its predecessor.[4] These comparisons need context: Meituan says the table combines publicly reported scores for some rival models with its own self-run results, which it describes as indicative only.[2]

Meituan says users can try LongCat-2.5 on its website and developers can call it through an API launched alongside the model.[1] LongCat’s API documentation lists support for agent harnesses including Claude Code, Codex and Hermes.[3] The API price table lists uncached input at $0.30, cached input at $0.006 and output at $1.20.[6] Meituan also reports that its W4A4 inference system quantizes expert weights and activations to 4 bits, with comparable performance to the unquantized model on all evaluated downstream tasks.[2] In online stress tests against a W8A8 INT8 baseline, the company reports higher prefill and decode throughput and lower TTFT and TPOT.[2]