LongCat-2.0 is a newly open-sourced, large-scale Mixture-of-Experts (MoE) language model featuring 1.6 trillion total parameters with approximately 48 billion activated per token. Notably, its entire training and deployment were executed entirely on AI ASIC superpods—spanning 35 trillion tokens without any irrecoverable loss spikes—demonstrating the viability of frontier-scale training on alternative hardware.
Architecturally, the model introduces LongCat Sparse Attention (LSA) for efficient 1M-token context processing, combined with an N-gram Embedding module to maximize parameter efficiency. Supported by a specialized multi-expert post-training pipeline, LongCat-2.0 excels in coding, reasoning, and autonomous agentic tasks.
Benchmark results highlight its highly competitive performance, rivaling top-tier proprietary models like GPT-5.5 and Claude Opus 4.x in evaluations such as Terminal-Bench and SWE-bench. Deeply integrated with mainstream frameworks like Claude Code, it offers developers a highly stable and efficient collaborative experience for complex repository-level edits and real-world agentic workflows.
