Physix Frontier · News Briefing Card (IT Home · Nov 9, 2026)

Meituan LongCat open-sources LongCat-2.5 multimodal model

KEY FACTS

  • Meituan LongCat officially open-sourced the LongCat-2.5 model on November 9.
  • The model uses the MIT license, permitting commercial inference deployment and further training.
  • The model has 1.6 trillion total parameters, with about 48 billion parameters activated per token.
  • The model supports a one-million-token context and targets long-horizon agent tasks.
  • Visual signals are represented as multi-level discrete tokens and processed uniformly by the backbone model.

KEY DATA

1.6 trillionTotal parameters
About 48 billionActivated parameters per token
One millionContext length

PHYSIX OBSERVATION

With 1.6 trillion parameters plus a million-token context, Meituan has pulled the threshold for long-horizon agents directly into the open-source camp, and the MIT license allowing commercial use is the biggest show of good faith. Two things are worth noting: vision and text share backbone parameters rather than sitting in an add-on module; and W4A4 quantization is clearly aimed at the memory and bandwidth shortfalls of domestic chips. Competition in open-source large models is shifting from parameter counts to end-to-end delivery and deployment cost.

Source: IT Home report