The Meituan LongCat team released the weights for LongCat-2.5 one week after its release. The license is licensed under the MIT terms and can be used commercially.

This model is designed for agents that are capable of independently executing long-duration tasks, with programming being the key focus. Each token activates approximately 48 billion parameters, and the model supports input in the form of text, images, and videos. On Terminal Bench v2.1, it scored 88.4 points, while Claude Opus 5 achieved a score of 89.1.

The team specifically emphasized that Huawei Ascend acceleration cards were used throughout the entire training process, and no NVIDIA GPUs were utilized.

On the very same day, an organization named Unemployment Labs released an analysis stating that the initial parameters in the public weights of six models—DeepSeek, Moon’s Dark Side, MiniMax, and Ant Group’s models—corresponded with the computational architecture of NVIDIA’s Blackwell chip. The results obtained using DeepSeek-V4-Flash were successfully replicated on a B300 hardware platform, while similar results were also achieved when using Kimi K3 on a GB200 device.

Their conclusion is that these models were at least initialized using the Blackwell method, and it is very likely that the relevant chips were also used in the training process. Such accusations are particularly sensitive in the context of export controls.

When considered together, these two factors become rather intriguing. In the past, it was extremely difficult to verify claims regarding the use of domestically produced hardware for training models; however, now some people have begun to look for evidence directly within the weights themselves. If this approach proves valid in the long run, anyone who claims that they “do not use NVIDIA technology” will likely be subjected to such verification – and LongCat-2.5 is probably no exception.