LongCat-2.5 at a glance
Meituan published the full weights of LongCat-2.5 on Hugging Face under the MIT license on November 9, 2026. The native multimodal mixture-of-experts model activates about 48 billion of its 1.6 trillion parameters per token and is built for long-horizon agent and coding work.
What changed with LongCat-2.5
Meituan made the full weights of LongCat-2.5 public and ungated on Hugging Face on November 9, 2026, under the permissive MIT license. LongCat-2.5 is a native multimodal mixture-of-experts model with 1.6 trillion total parameters and roughly 48 billion activated per token, built on LongCat-2.0 for long-horizon agentic tasks: it combines visual understanding with agentic coding for codebase development, full-project builds, creative design and 3D work. The weights release follows the model's November 2 debut on Meituan's API, which launched at a limited-time $0.30 per million input tokens and $1.20 per million output tokens.
Meituan's model card reports benchmark results measured in-house or cited from public reports: 88.4 on Terminal-Bench v2.1 and 91.9 on BrowseComp, 63.1 on Humanity's Last Exam with tools, 83.1 on MCP Atlas and 64.6 on DeepSWE, alongside visual results such as 78.6 on Vision2Web. Meituan flags the unmarked scores as indicative, and Artificial Analysis has not scored the model, so no independent intelligence ranking exists yet. The card also describes W4A4 quantization of expert weights for native INT4 computation on domestic AI accelerators, hybrid thinking and non-thinking modes, and evaluations run at sequence lengths up to 1 million tokens.
LongCat-2.5 in numbers
Values are copied from the linked source on the date shown. They are not IA Compa measurements.
- Terminal-Bench v2.1 (Meituan model card)as of · Source for Terminal-Bench v2.1 (Meituan model card): huggingface.co (opens in a new tab)
- 88.4
- BrowseComp (Meituan model card)as of · Source for BrowseComp (Meituan model card): huggingface.co (opens in a new tab)
- 91.9
Where to compare LongCat-2.5
Independent leaderboards and official resources for comparing this model yourself. Rankings change often; each site shows its own current data.
Why LongCat-2.5 matters for businesses
Why it matters: a frontier-scale multimodal agent model that any business can download, self-host and use commercially under MIT changes the build-versus-buy math for agent products, at least for teams that can feed a 1.6-trillion-parameter model. For agencies, open weights at this scale mean client deployments no longer have to rent every token from a closed API — though the hardware bill keeps that option realistic only for larger clients, and the self-reported benchmarks still need independent confirmation.
Get AI News by email
Verified AI stories, each with its official source, on Monday and Thursday mornings. Add breaking news to hear about the biggest ones as soon as we publish them.
Every story links its official source
Free, in English or Spanish
Unsubscribe with one click, any time
Sources for this story
- LongCat-2.5 model card and weights — Hugging Face (opens in a new tab)Official
Meituan LongCat · huggingface.co
Last verified: . IA Compa summarizes public announcements for business readers. We are not affiliated with the companies mentioned. Always confirm details, pricing, and availability with the official source.
- #longcat
- #open-weights
- #mit-license
- #mixture-of-experts
- #ai-agents
- #multimodal
