Meituan releases LongCat-2.5, a 1.6T-parameter agent model trained on Chinese chips, API only for now

Meituan released LongCat-2.5 on November 2. It's the final version of LongCat-2.5-Preview, which went live on September 25, and it builds on LongCat-2.0 from June. It's a mixture-of-experts model with 1.6 trillion total parameters and about 48 billion active per token, aimed at long agent tasks and coding.

Two changes stand out in Meituan's write-up. Images become discrete tokens that the same model reads alongside text, with no separate vision encoder, which Meituan uses for rebuilding web pages from mockups and building 3D pages. And the agent training leaned on long trajectories: mid-training sequences of up to 1M tokens, large-scale reinforcement learning, and failures from real agent runs turned back into reproducible training tasks. Meituan says the whole training run used Chinese accelerators, and it serves the model with 4-bit weights and activations.

All scores are Meituan's, and some competitor numbers were measured by Meituan. Per Neural Brew's reading of the benchmark table, it gets 88.4 on Terminal Bench v2.1 (Opus 5: 89.1) but 38.5 on Terminal Bench v4.0 (Opus 5: 53.9). The comparisons stop at Opus 5 and GPT-5.6 Sol, not the newest models.

It's on longcat.ai and through Meituan's API. Neural Brew reports promotional pricing of $0.30 per million uncached input tokens, $0.006 cached and $1.20 output, and says the API changelog lists Claude Code, Codex and Hermes as compatible harnesses. Weights are the open question: Neural Brew found no announcement, while TechNode (citing IT Home) says GitHub and Hugging Face releases are planned.