LongCat-2.5

Meituan LongCat has open-sourced LongCat-2.5, a natively multimodal model with 1.6 trillion total parameters, roughly 48 billion activated per token, and one-million-token context, released under the MIT license for commercial inference and retraining. It encodes visual signals as multi-level discrete tokens sharing the same backbone as text, and uses a W4A4 4-bit quantization scheme tuned for domestic AI chips constrained by memory and bandwidth.

OpenCode announced that LongCat 2.5 is free on its platform for two weeks, highlighting a 1M-token context window, multimodal input and zero data retention. The move is a distribution play on a third-party coding tool, with no model size, benchmark results or pricing details disclosed.