Meituan Releases LongCat-2.5, a 1.6 Trillion-Parameter Multimodal AI Model for Agents
Reported by lapaasvoice.com, Hacker News and Tech Times3 sources ↓
Meituan released LongCat-2.5, a multimodal model for long-running agent tasks. It is available through an API and the LongCat website; its weights have not been publicly released.
Key points
- The model activates about 48 billion of its 1.6 trillion parameters per token
- Developers can access it through Meituan’s API and users can try it on the LongCat website
- Meituan says it deployed the model in its Miaoshow merchant-management workbench
- Meituan’s performance comparisons are based on in-house tests; no independent evaluation is reported
- Model weights are not public; Meituan says GitHub and Hugging Face releases are expected to follow
Sources
Details
LongCat-2.5 is a mixture-of-experts model with 1.6 trillion total parameters and about 48 billion activated per token. Meituan says it was trained on domestic computing infrastructure and is compatible with agent tools including Claude Code, Codex and Hermes. The model adds native image understanding to capabilities for coding and multi-step tasks.
Tech Times reports that Meituan deployed the model in Miaoshow, a workbench for restaurant and hotel operators to analyze business performance and carry out operational tasks. Meituan’s claims that the model performs between Claude Opus 4.8 and Claude Opus 5 on several evaluations, and exceeds Claude Opus 5 on some tasks, have not been independently assessed, according to Tech Times.
Written from coverage as of
Images
- lapaasvoice.com
- Tech Times