The model, and what it is for
Meituan announced LongCat-2.5 on 2 November, describing it in its release post as “a native multimodal model for long-horizon agentic tasks” with 1.6 trillion total parameters and approximately 48 billion activated per token. It builds on LongCat-2.0, and the company says the gains came from folding more long-horizon trajectories into long-sequence training and running large-scale reinforcement learning on end-to-end engineering work — complex codebase development and full-project builds.
One detail of the training pipeline is worth naming: Meituan says it converts real-world agent failure trajectories into reproducible, verifiable training tasks, and filters GitHub pull requests and linked issues into synthetic environments validated with fail-to-pass and pass-to-pass tests.
On the multimodal side, the model represents visual signals as multi-level discrete tokens, so the backbone handles images and text in one token format rather than through a separate vision pathway.
The company’s own scores
Meituan published an evaluation table, and every number in it is vendor-reported. The sharpest line is Terminal Bench v4.0, where LongCat-2.5 scores 38.5 against 3.0 for LongCat-2.0 — a jump of more than twelve times in one release. On the same row Meituan lists Claude Opus 5 at 53.9, GLM-5.3 at 41.8 and GPT-5.6-Sol at 37.3, so the model does not lead the benchmark it improved most on.
The pattern repeats across the agentic coding rows. On DeepSWE v1.1 the model goes from 9.7 to 64.6, with Claude Opus 5 at 68.8 and DeepSeek-V4.1-Flash at 74.2. On Program Bench it moves from 2.5 to 25.5, against 37.0 for Claude Opus 5. Where Meituan’s table does put it first is browsing and general agent work: 91.9 on BrowseComp, ahead of the 91.5 and 91.2 it lists for DeepSeek-V4.1-Flash and kimi-k3, and 81.0 on ResearchRubrics. None of these figures has been independently reproduced.
Chinese silicon, and no weights yet
The infrastructure section carries the line most likely to travel. Meituan says it optimised long-context training and W4A4 inference “on domestic AI accelerators” — that is, Chinese-made chips rather than Nvidia hardware — and that numerical consistency between training and inference plus long-context rollout work made the large-scale reinforcement learning practical.
The model is available to try through Meituan’s chat interface and its API platform. The GitHub and Hugging Face links on the release page both read “Coming soon,” so the weights are not out, and the only scores anyone can check are the ones in Meituan’s table.
What to watch
Two things would settle this. An independent evaluation on the agentic coding benchmarks would show whether the twelvefold Terminal Bench jump holds outside the vendor’s harness. And a weights release would let anyone test the claim about domestic accelerators by running the model somewhere else.
