Meituan · 2026-11-10 · major
LongCat-2.5 — Meituan opens weights of its 1.6T agentic model
In short
Meituan released the weights of LongCat-2.5 under the MIT license. It is a 1.6T-parameter multimodal mixture-of-experts model with about 48B active per token, built for long-horizon agent tasks.
Meituan · 2026-11-10 · 3 sources checked
Meituan's LongCat-2.5 puts a 1.6T multimodal agent model on Hugging Face under the MIT license.
Quick facts
| Maker | Meituan (LongCat team) |
|---|---|
| License | MIT (weights) |
| Size | 1.6T total, ~48B active per token |
| Modes | Hybrid thinking and non-thinking |
| Terminal Bench v2.1 | 88.4 (self-reported) |
| Availability | Hugging Face weights, longcat.ai chat |
Benchmarks
What is it?
LongCat-2.5 is Meituan's native multimodal mixture-of-experts model for long-horizon agentic tasks. The weights are now on Hugging Face under the MIT license, after Meituan unveiled the model on November 1. It succeeds LongCat-2.0.
How does it work?
The model has 1.6 trillion total parameters and about 48 billion active per token, with hybrid thinking and non-thinking modes. The model card says the expert weights are quantized to W4A4.
Why does it matter?
Open weights at this size with a permissive license give teams a self-hostable agent model that Meituan's own benchmarks place close to closed frontier models. The scores are self-reported, so independent tests still need to confirm them.
Who is it for?
agent builders, teams that self-host large models
Frequently asked questions
- Is LongCat-2.5 open source and free for commercial use?
- LongCat-2.5 weights are published on Hugging Face under the MIT license, which allows commercial use and fine-tuning. The model card adds that the license does not grant rights to Meituan trademarks or patents. Teams should read the LICENSE file in the repository before shipping a product.
- How does LongCat-2.5 compare with Claude Opus 5 and GPT-5.6-Sol?
- On Terminal Bench v2.1, Meituan's model card reports 88.4 for LongCat-2.5, against 89.1 for Claude-Opus-5 and 88.8 for GPT-5.6-Sol. Many LongCat scores are in-house measurements, and the card says they are indicative only. No independent evaluation was found.
- What is new compared with LongCat-2.0?
- LongCat-2.5 is the successor to LongCat-2.0 and is a native multimodal model with hybrid thinking and non-thinking modes. It keeps the 1.6T-parameter mixture-of-experts layout with about 48B active parameters per token. The model card lists the expert weights as quantized to W4A4.
- Where can LongCat-2.5 run?
- The LongCat-2.5 model card points to an NPU deployment guide built on SGLang-FluentLLM. No Hugging Face inference provider hosted the model at the time of checking. A chat interface at longcat.ai lets people try it without downloading weights.
Try it
https://huggingface.co/meituan-longcat/LongCat-2.5