All news

Meituan releases multimodal LongCat-2.5 weights with 1.6 trillion parameters under the MIT License

Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .

A model with trillions of total parameters can now be tested and modified in a private environment without hosted API restrictions, although full deployment still requires infrastructure with terabytes of memory.

Meituan has published the full LongCat-2.5 weights on Hugging Face under the MIT License, one week after launching the hosted API. The multimodal MoE model contains 1.6 trillion parameters, approximately 48 billion of which are active per token. It accepts text, images and video and is designed for long-horizon agent tasks involving codebases, complete projects, design and 3D. The model checkpoint occupies about 1.8 TB, so local deployment requires the full weight set even though only some experts are active. In Meituan's own measurements, LongCat-2.5 scored 88.4 on Terminal Bench v2.1 versus 70.8 for LongCat-2.0 and 64.6 on DeepSWE v1.1 versus 9.7. VentureBeat notes that comparisons with newer proprietary models on Terminal Bench v4.0 no longer favor it and that the test settings differ. The license permits commercial use, modification and redistribution with the required notice. W4A4 quantization of experts and activations is available for Chinese accelerators; Meituan's own stress tests show prefill throughput increasing by 43.3% and decode throughput by 42.77% relative to W8A8.

History

Meituan releases LongCat-2.5

3 events
  1. Meituan releases multimodal LongCat-2.5 weights with 1.6 trillion parameters under the MIT LicenseOpen now
  2. Meituan releases multimodal LongCat-2.5 with 1.6 trillion parameters
  3. Meituan releases LongCat-2.5-Preview, a 1.6 trillion-parameter model for long agentic tasks at $0.30 per million tokens