LongCat-2.0 is a large-scale sparse Mixture-of-Experts language model developed by Meituan, activating roughly 48 billion parameters per token out of 1.6 trillion total. To preserve quality at long horizons, the team introduced LongCat Sparse Attention and trained the model on hundreds of billions of tokens of 1M-context data, giving it a native 1,000,000-token context window. Both pretraining and large-scale deployment were carried out entirely on AI ASIC superpods, spanning millions of accelerator-days across more than 35 trillion tokens without rollbacks or irrecoverable loss spikes, which positions the release as a demonstration of frontier-scale training on alternative accelerator hardware.

The model is positioned for coding, repository-level edits, and agentic workloads, with deep integration into mainstream developer harnesses such as Claude Code, OpenClaw, and Hermes. Benchmark coverage called out by the announcement spans Terminal-Bench 2.1, SWE-bench Pro, SWE-bench Multilingual, FORTE, RWSearch, and BrowseComp, indicating attention to terminal-style automation, software engineering, multilingual code, function calling, real-world search, and web browsing tasks. Practically, this makes LongCat-2.0 a strong fit for teams that need long-context reasoning over large codebases or multi-step agentic pipelines, especially when the cataloged API limit and open weights published under the meituan-longcat organization on Hugging Face are required.