Meituan's LongCat team has published the weights of LongCat-2.5 on Hugging Face under the MIT license. LongCat-2.5 is a 1.6-trillion-parameter mixture-of-experts model that the model card describes as a "native multimodal MoE model for long-horizon agentic tasks." About 48 billion parameters are active for each token.
The model takes images and text as input and produces text. It turns native-resolution visual features into discrete tokens. Thinking can be switched on or off with an enable_thinking flag. The MIT license covers the model weights; the model card says Meituan's trademarks and patents are not licensed.
On Meituan's own reported benchmarks, LongCat-2.5 scores 88.4 on Terminal Bench v2.1, 64.6 on DeepSWE v1.1, 91.9 on BrowseComp, 83.1 on MCP Atlas and 63.1 on Humanity's Last Exam with tools. The card lists comparison scores from public sources that put it near current closed frontier models. It is ahead of the cited Claude Opus 5 and GPT-5.5 scores on BrowseComp, and slightly behind the cited Claude Opus 5 scores on Terminal Bench, MCP Atlas and HLE. Meituan says scores without public citations come from its own evaluations and are indicative only.
The model is built for Chinese AI accelerators. Expert weights use W4A4 quantization, with 4-bit weights and activations computed natively in INT4, and deployment is supported on NPU platforms through SGLang-FluentLLM. Against a W8A8 baseline, Meituan reports 43.3% higher prefill throughput and 42.77% higher decode throughput.
LongCat-2.5 is one of the largest open-weight models available under a fully permissive license. Companies can deploy or fine-tune it commercially without separate permission. At this size, though, self-hosting needs a multi-node cluster, so most developers will likely use it through hosted inference providers rather than on their own hardware.