Meituan Open-Sources 1.6T Parameter Multimodal Model LongCat-2.5 Under MIT License

Meituan's LongCat team officially open-sourced its native multimodal large model, LongCat-2.5, on the evening of November 9. The model features 1.6 trillion total parameters, activates approximately 48 billion parameters per token, and supports a context window of 1 million tokens. Released under the permissive MIT license, it allows for commercial inference deployment and secondary training without additional authorization requirements.

Technically, LongCat-2.5 focuses on optimizing visual programming, complex codebase development, and long-horizon agent tasks. To address the limitation of domestic AI accelerators having less than 80GB of VRAM, Meituan provided NPU deployment instructions based on the SGLang-FluentLLM project. The release includes a W4A4 quantized version, which quantizes expert weights and activations to 4 bits. According to official data, this version improves prefill and decode throughput by 43.3% and 42.77% respectively compared to the W8A8 INT8 baseline.

In benchmark comparisons against eight other major models including Claude Opus 5 and GPT-5.6-Sol, LongCat-2.5 ranked first in four specific tests: BrowseComp (web retrieval), ALE v1.1 (cross-domain agents), ResearchRubrics (deep research), and 3DCodeBench (text-to-3D code generation). However, its performance in some general coding tests still trails behind competitors like DeepSeek-V4.1-Flash.

Regarding API pricing, LongCat-2.5 currently offers a limited-time rate of 2 RMB per million input tokens and 8 RMB per million output tokens, significantly lower than comparable overseas models, though its price advantage over top domestic models is less pronounced. Early developer feedback indicates strong capabilities in generating interactive 3D scenes and game frontends, but also notes issues with circular reasoning that can reduce execution efficiency.