LongCat-2.5 weights now on Hugging Face under MIT license
Meituan's 1.6T agentic model is now downloadable, with no usage caps in the license.
Meituan's LongCat team has published the full LongCat-2.5 weights on Hugging Face under the MIT License. As of November 10, 2026, the repository meituan-longcat/LongCat-2.5 is public and ungated, next to a separate 4-bit meituan-longcat/LongCat-2.5-W4A4 repository. The model launched on the LongCat API on November 2 without weights (our launch post).
What changed
- Weights:
meituan-longcat/LongCat-2.5, 208 safetensors shards, with a native multimodal mixture-of-experts design of 1.6 trillion total parameters and about 48 billion active per token. - License: MIT for the model weights. The model card says it grants no rights to Meituan trademarks or patents.
- Quantized version:
meituan-longcat/LongCat-2.5-W4A4applies 4-bit quantization to expert weights and activations for native INT4 compute on domestic AI accelerators. It is also MIT. - Serving: the model card documents NPU deployment through SGLang-FluentLLM. It gives no GPU serving recipe yet.
- Thinking is a switch:
enable_thinkingturns reasoning on or off through the chat template.
What it means
Any host can now serve LongCat-2.5 without a separate license agreement, and teams can run it on their own hardware. LongCat's own API still lists LongCat-2.5 with a 1M-token context window.
On RouterPlus
LongCat-2.5 is not in the RouterPlus catalog as of November 10, 2026. Compare current open-weight options on the RouterPlus models page.