The release and terms of use
According to a report by IT之家, Meituan announced on November 9 that it was open-sourcing LongCat-2.5, the multimodal model unveiled on November 2 and designed for agent tasks requiring prolonged sessions. The release is under the MIT license: the source specifies that the license allows deployment for commercial inference and additional training without further authorization.
The move from launch to open source took one week, a short interval that signals the intention to make the model quickly available to developers and enterprises. The actual performance claimed and the completeness of the published materials remain to be independently verified; the source says they are distributed via Hugging Face, a technical blog, and an API platform.
Architecture and stated figures
The specification sheet provided by the source indicates a remarkable scale: 1.6 trillion total parameters, with about 48 billion activated per token, and a context window of up to one million tokens. These values place the model in the large-system tier, but the source does not report direct comparisons with other models or detailed benchmark results.
Meituan claims improvements in end-to-end delivery capability in scenarios such as development on complex codebases, building complete projects, Visual Coding, and professional office work. These are company claims: the material contains no independent measurements to support them.
The bet on visuals and domestic chips
A central technical element concerns image processing. According to the source, the model represents visual signals as discrete tokens at multiple levels, handled by the same base model together with text, instead of relying on a separate visual module. The source describes residual vector quantization on native-resolution features, with visual tokens that can be precomputed and stored offline to reduce online computation and balance loads in distributed training; the visual vocabulary is reportedly trained jointly with the base model, sharing parameters and semantic space between visual understanding and code generation.
The source also notes a specific focus on domestic compute chips with memory and bandwidth constraints: the company says it worked in a coordinated way on architecture, hardware adaptation, and deployment strategies. On the inference side, LongCat-2.5 adopts a W4A4 scheme that quantizes both expert weights and activations to 4 bits, performing computations through native INT4 units.
These details are relevant because they suggest a model designed not only for scale but also for efficiency on alternative hardware. However, the source does not quantify gains in terms of memory, speed, or cost, nor does it clarify which chips were actually tested: these are aspects that require verification against the published technical materials.
