LongCat-2.5
Meituan released LongCat-2.5, a 1.6 trillion parameter native multimodal model with about 48 billion active parameters per token, built for long-horizon agentic coding.
It builds on LongCat-2.0 and adds native visual understanding by representing images as multi-level discrete tokens in the same stream as text. Training added more long-horizon trajectory data and large-scale long-horizon reinforcement learning, and failure trajectories from real agents are turned into verifiable training tasks. Meituan also optimized W4A4 inference (4-bit weights and activations) on domestic chips.
- Date
- Monday 2 November 2026
- Lab
- Meituan
- Kind
- model
- Access
- closed API
Figures
| Measure | Value | Measured by |
|---|---|---|
| Total parameters | 1.6T about 48B active per token | company |
| Context window | 1M tokens output up to 128K tokens, per Crypto Briefing | company |
| Terminal Bench v2.1 | 88.4 LongCat-2.0 scored 70.8 | company |
| AutomationBench | 51.8 LongCat-2.0 scored 15.8 | company |
| BrowseComp | 91.9 | company |
| PaperBench | 82.1 | company |
| 3DCodeBench | 55.4 | company |
| Vision2Web | 78.6 | company |
Available at longcat.chat and via API; Meituan says GitHub and Hugging Face releases are coming soon. Benchmark figures are company-reported and not independently verified. It follows a preview on 2026-09-25.
Sources
- pandaily.com/meituan-longcat-2-5-official-release-1-6t-native-multimodal-long-horizon-agen
- technode.com/2026/11/02/meituan-launches-longcat-2-5-multimodal-ai-model-for-extended-codi
- cryptobriefing.com/meituan-releases-longcat-2-5-multimodal-model/
This record was checked against its sources on 2 November 2026. How we check