正文 · AI 翻译
IT Home reported on November 9 that Meituan LongCat released LongCat-2.5, a native multimodal model for long-horizon agent tasks, on November 2. Today (the 9th), Meituan announced that the model is officially open source.
The model is released under the MIT license, supports commercial inference deployment and further training without additional authorization, and is a native multimodal model for long-horizon agent tasks. Its end-to-end delivery capabilities have improved significantly in complex codebase development, complete project construction, Visual Coding, and professional office scenarios.
The model has 1.6 trillion parameters in total, activates approximately 48 billion parameters per token, and supports a context length of 1 million tokens. For domestically produced computing chips constrained by memory capacity and bandwidth, the team also carried out extensive coordinated optimization of the model architecture, chip adaptation, and deployment strategy.
According to the introduction, LongCat-2.5 performs well in Visual Coding and supports tasks such as creative design and 3D work. These improvements stem from changes to the visual representation layer: the model represents visual signals as multilevel discrete tokens, which are processed by the backbone model in the same way as text, rather than by an add-on visual module. Specifically, the model applies residual vector quantization to visual features at their native resolution, progressively filling in details; visual tokens can be precomputed and cached offline, reducing online computation and making distributed training workloads more balanced; and the visual codebook is trained jointly with the backbone, with vision and text sharing the same backbone parameters to collaboratively perform visual understanding and code generation in the same semantic space.
In addition, LongCat-2.5 uses an efficient W4A4 inference scheme that quantizes both expert weights and activations to 4 bits and performs computations using native INT4 matrix multiplication units.
IT Home provides the relevant links below:
Open-source model: https://huggingface.co/meituan-longcat/LongCat-2.5
Technical blog: https://longcat.ai/blog/longcat-2.5/
API platform: https://longcat.chat/platform/product
Advertising disclosure: External links in this article (including but not limited to hyperlinks, QR codes, and access codes) are provided to convey more information and save time in selecting resources. The results are for reference only. All IT Home articles containing external links include this disclosure.