Meituan released LongCat-2.5 on November 2, a native multimodal model with 1.6 trillion total parameters built for long-horizon agent work such as engineering inside large codebases and building complete software projects. The model activates about 48 billion parameters per token and extends LongCat-2.0 with stronger visual understanding and agentic coding.
LongCat-2.5 is available now through Meituan's online chat service and its API platform. Meituan said releases on GitHub and Hugging Face will follow.
Meituan evaluated the model on nine benchmarks covering code, general tasks and visual agents. LongCat-2.5 scored 91.9 on BrowseComp, which Meituan ranked first, and 51.8 on AutomationBench, ahead of Anthropic's Claude Opus 5 at 50.3. On 3DCodeBench, a text-to-3D project test, it scored 55.4, a narrow lead over the comparison set. It also led on LongCatClawBench, though no score for that test was published in coverage of the release.
Key facts
- 01Meituan released LongCat-2.5 on November 2, 2026, a native multimodal model built for long-horizon agent tasks.
- 02LongCat-2.5 has 1.6 trillion total parameters, with about 48 billion activated per token.
- 03Meituan reported 91.9 on BrowseComp and 51.8 on AutomationBench, where Claude Opus 5 scored 50.3.
- 04The model is live in Meituan's chat service and API platform, with GitHub and Hugging Face releases planned.
- 05Meituan optimized long-context training and W4A4 inference on domestic Chinese compute.
The comparison set included OpenAI's GPT-5.5 and GPT-5.6-Sol alongside Claude Opus 4.8 and Claude Opus 5. Chinese coverage of the results noted that LongCat-2.5 still trails the top overseas models somewhat on code and terminal-command tasks. All of the scores are Meituan's own reported numbers and have not been independently verified.
The architectural change is in how the model sees. Meituan represents visual signals as multi-level discrete tokens, so the same backbone processes images and text in one unified token format. Meituan says that design supports visual coding, creative design and 3D tasks.
On training, Meituan added more long-trajectory data to its long-sequence training and ran large-scale, long-horizon reinforcement learning aimed at end-to-end engineering on complex codebases and full project builds. The company also converts failures from real agent runs into reproducible, verifiable training tasks.
The infrastructure work ran on domestic Chinese compute. Meituan said it optimized long-context training and W4A4 inference, meaning 4-bit weights and 4-bit activations, and used training-inference numerical consistency and long-context rollout optimization to support reinforcement learning at scale. Meituan did not name the chip vendors involved.
Reports on the release did not include pricing or a context-window figure for LongCat-2.5. The GitHub and Hugging Face releases are the next step Meituan has announced.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
