LongCat-Flash

LongCat-Flash is Meituan's first open-source large language model, a 560-billion-parameter mixture-of-experts model that activates roughly 27B parameters per token. Released in September 2025 with Chat and Thinking variants, it uses multi-head latent attention and a dynamic-activation MoE design that Meituan reports serving at over 100 tokens per second on H800-class hardware. Both weights and code are MIT licensed, unusually permissive for a model of this scale.

The release marked Meituan's pivot from delivery super-app to AI infrastructure player, followed by the 1.6T LongCat-2.0 in 2026. Benchmark claims, including strong coding results for the Thinking variant, are vendor-reported and should be treated as directional until independently reproduced. Meituan also offers generous free API tiers, but the real story is the weights: few MIT-licensed models operate at this scale.