Releases from frontier labs, major companies and research institutes — grouped by model, deduplicated across HuggingFace, OpenRouter and official blogs.
Click a week to filter
Meituan LongCat🇨🇳Major lab
LongCat-Flash-Lite-Sparse is a non-thinking Mixture-of-Experts (MoE) model with 69B total parameters and approximately 3B activated parameters per token. Built on LongCat-Flash-Lite, it replaces dense MLA with LongCat Sparse Attention (LSA) and natively supports context lengths of up to 1M tokens.
Meituan LongCat🇨🇳Major lab
LongCat 2.0 is a sparse mixture-of-experts language model from Meituan, with 48B active parameters out of 1.6T total. It is suited for coding, repository-level changes, long-horizon problem solving, and agentic...