LongCat-2.0 weights are out, with clear agent cost angles
LongCat-2.0 is now available on Hugging Face in INT8 and FP8 weight versions. The model is described as having 1.6 trillion total parameters, but only about 48 billion are active for each token it processes. It was trained with large amounts of 1-million-token context data and is positioned for coding, repository-level edits, tool use, and other AI agent work.
Reported scores include 70.8 on , 59.5 on SWE-bench Pro, 77.3 on SWE-bench Multilingual, and 73.2 on FORTE, alongside comparisons with major . It can be deployed with SGLang, but the recommended GPU setup uses 16 H20 cards, so this is not a small local model for ordinary hardware. Its chat template includes tool-calling examples and options to turn thinking mode on or off, with thinking mode off presented as a way to improve .
The weights use the MIT license, while the model card still warns teams to test accuracy, safety, fairness, and language performance before using it in sensitive settings.
Key points
- LongCat-2.0 INT8 and FP8 weights are available on Hugging Face.
- It has 1.6 trillion total parameters, with about 48 billion active per token.
- The model targets coding, long-context work, tool use, and s.
- SGLang deployment is supported, but the suggested setup needs 16 H20 GPUs.
- The chat template includes a thinking mode switch that can help test .