Models
Meituan publishes the LongCat-2.5 weights on Hugging Face under the MIT license: a 1.6 trillion-parameter mixture-of-experts model with about 48 billion active per token; benchmark scores are the vendor’s own
Meituan released LongCat-2.5 weights under the MIT license: 1.6T parameters, about 48B active. Benchmarks are the vendor’s; “open source” means open weights.
Meituan put the full weights of LongCat-2.5 on Hugging Face on November 9, 2026, under the MIT license, about a week after announcing the model. It is a 1.6 trillion-parameter mixture-of-experts model with about 48 billion parameters active per token. The headline is open weights, which is narrower than fully open source, and the benchmark scores come from Meituan.
The essentials
- What is open: AI Weekly and Superpower Daily report the weights on Hugging Face under the MIT license, allowing commercial use and modification. AI-TLDR adds that the model card says the license grants no rights to Meituan trademarks or patents. We found no mention of released training data.
- Size and design: about 1.6 trillion total parameters, 48 billion active per token, a 1M-token context, vision built in. Superpower Daily notes Hugging Face metadata lists 1.8 trillion, and AI Weekly puts the checkpoint at 1.8 TB.
- Claimed performance: AI Weekly cites 88.4 on Terminal Bench v2.1 and 64.6 on DeepSWE v1.1, up from 70.8 and 9.7 for LongCat-2.0. These are Meituan’s numbers as relayed; we found no independent replication.
- Timeline: a preview appeared in late September, the full model was announced around November 1, and an earlier report from Startup Fortune still said weights were not published. Hosted pricing is reported as $0.30 per million uncached input tokens and $1.20 output.
Open weights, not full open source
MIT-licensed weights let anyone run, modify and redistribute the model. That is not the same as open training data or code, which we did not find released. Running 1.8 TB of weights also needs serious hardware.
Why the numbers need care
The jump from 9.7 to 64.6 on one benchmark is large and self-reported. Parameter counts differ by source, 1.6 versus 1.8 trillion. Not investment advice.
What is not established
- Independent replication of the benchmark scores: none found.
- Whether training data or full training code was released: not mentioned in the sources.
- Why Hugging Face metadata lists 1.8 trillion against Meituan’s 1.6 trillion: not explained.
- Compliance terms for US companies using Chinese MIT-licensed weights: not addressed by the sources.
Summary table
| Point | What is stated as of November 7, 2026 | Source |
|---|---|---|
| Release | Weights on Hugging Face on November 9, 2026, model announced around November 1 | AI-TLDR |
| License | MIT, with no rights to Meituan trademarks or patents per the model card | AI-TLDR |
| Size | 1.6 trillion parameters, about 48 billion active per token | AI Weekly |
| Benchmarks | 88.4 on Terminal Bench v2.1 and 64.6 on DeepSWE v1.1, vendor-reported | AI Weekly |
| Earlier status | Preview in September; weights not yet published at the time of one report | Startup Fortune |
FAQ
Is LongCat-2.5 open source?
Its weights are public under the MIT license. We found no sign that training data was released, so open weights is the safer term.
How big is it?
About 1.6 trillion parameters in total with roughly 48 billion active per token; Hugging Face metadata lists 1.8 trillion.
Are the benchmark scores independent?
No. They are Meituan’s figures as relayed, and we found no independent replication.
Method and limits
Sources checked on November 10, 2026: AI Weekly, Superpower Daily, AI-TLDR, Cocoloop and Startup Fortune, read as search extracts; the Hugging Face model card was not read directly. Not investment advice.
Sources
- AI Weekly: Meituan Open-Sources LongCat-2.5, 1.6T MoE, Under MIT License (verified 2026-11-08)
- Superpower Daily: Meituan publishes LongCat-2.5 weights, opening a self-hosted route for visual coding (verified 2026-11-08)
- AI-TLDR: LongCat-2.5 — Meituan opens weights of its 1.6T agentic model (verified 2026-11-08)
- Cocoloop: Meituan Launches LongCat-2.5-Preview, Targets Long-Horizon Tasks (verified 2026-11-08)
- Startup Fortune: Meituan's LongCat-2.5 Jumps Into China's Open AI Coding Race (verified 2026-11-08)
Spotted an error in this article? Report it · Our corrections policy