Meituan Releases LongCat-2.5 Weights Under MIT: the 1.6T-Parameter, ~48B-Active Native Multimodal MoE Lands on Hugging Face With a W4A4 Quantized Version
On the evening of November 9 SGT, the official Meituan LongCat account announced that LongCat-2.5 weights are now open under the MIT license on Hugging Face. LongCat-2.5 is the native multimodal MoE model Meituan launched as an API on November 2, with 1.6 trillion total parameters and about 48 billion active per token, aimed at long-horizon agent tasks, agentic coding, creative design and 3D work. Meituan published both the BF16 main model (a repository of about 3.56TB) and a W4A4 quantized version (about 1.34TB). The model card currently only documents NPU deployment, via SGLang-FluentLLM, and the model supports both thinking and non-thinking modes. Vendor-reported scores include 88.4 on Terminal Bench v2.1 and 91.9 on BrowseComp; the model card says LongCat scores are measured in-house and are indicative only.
Key Takeaways
- Open weights: LongCat-2.5 is on Hugging Face under the MIT license (main model, W4A4 version), one week after the November 2 API launch
- Scale: a native multimodal MoE with 1.6T total parameters and about 48B active; the BF16 repository is about 3.56TB and the W4A4 repository about 1.34TB
- Deployment: the model card only documents NPU deployment for now (a SGLang-FluentLLM branch), and training and serving were done on domestic AI accelerators; running it on ordinary GPUs at home is not realistic
- Scores (vendor-reported, LongCat numbers measured in-house): Terminal Bench v2.1 88.4 (Claude Opus 5 at 89.1), BrowseComp 91.9, AutomationBench 51.8, DeepSWE v1.1 64.6
- Usage: a hybrid model with a thinking mode (keeping all reasoning content is recommended) and a non-thinking mode for better token efficiency, called through an OpenAI-style API with tool calling
