69B total / ~3B active per token (MoE)
Up to 1M tokens
MIT
Non-thinking open-weight Mixture-of-Experts language model built on LongCat-Flash-Lite, replacing its dense Multi-head Latent Attention with LongCat Sparse Attention and repeating the long-context extension stage on an enriched corpus. LongCat Sparse Attention combines streaming-aware indexing, which reserves fixed sink and local-window selections for more contiguous KV-cache access; cross-layer indexing, which reuses one indexing pass across adjacent layers; and optional hierarchical indexing, which narrows candidates through coarse-to-fine scoring. The model contains 69 billion total parameters while activating approximately 3 billion per token and natively supports context lengths up to one million tokens. Meituan reports substantial gains over the dense LongCat-Flash-Lite predecessor on agentic coding, search, and tool-use evaluations, including 68.2 on SWE-bench Verified without hierarchical indexing versus 54.4 for the dense model, and 95.18 on tau2 Telecom versus 72.8. The release provides approximately 138 GB of BF16 Safetensors across 26 shards, a Transformers chat template, and an SGLang deployment path capable of serving the model on a single H20 141 GB node. Hierarchical indexing can be enabled at inference time without additional training, allowing deployments to trade some benchmark quality for lower indexing overhead.