LongCat 2.0

LongCat 2.0 is a large-scale Mixture-of-Experts (MoE) language model developed by the LongCat team. With a total of 1.6 trillion parameters, it is architected for agentic coding, repository-level code understanding, and complex autonomous workflows. The model utilizes a sparse design where approximately 48 billion parameters (with a dynamic range of 33B to 56B) are activated per token, striking a balance between high-capacity reasoning and inference efficiency.

A core technical feature of LongCat 2.0 is LongCat Sparse Attention (LSA), which facilitates a native context window of 1 million tokens. LSA employs streaming-aware and cross-layer indexing to reduce the quadratic computational burden of standard attention, allowing the model to maintain state and coherence across extremely long sequences. Additionally, the model incorporates an N-gram Embedding module that significantly expands the embedding space to capture more granular local context and improve token-level representations.

The model was pretrained on over 35 trillion tokens of multilingual and programming data. Its training run is notable for being executed entirely on a 50,000-card cluster of domestic AI ASIC superpods, demonstrating stable frontier-scale pretraining independent of traditional GPU hardware. This process involved specialized techniques such as Zero-computation experts and a PID-controlled expert routing system to maintain consistent throughput and hardware utilization.

Optimized for agentic tasks, LongCat 2.0 supports native tool calling and multi-step reasoning. It is designed to function as a backbone for autonomous agents, excelling in code generation, refactoring, and automated debugging. The model's post-training involves a specialized multi-expert distillation process (MOPD) that fuses distinct agent and coding capabilities into the final 1.6T foundation.