Note that until last generation all Chinese models were relatively small. This was mainly due to lack of training hardware, but as soon as the new Huawei NPUs started shipping some Chinese labs switched to larger models:
Deepseek: 670B to 1.6T
Moonshot: 1.1T to 2.8T
Xiaomi: 310B to 1T
The new models also use a different architecture so I would assume in tests they will look different from previous generations.
Z.ai (GLM) and MiniMax on the other hand have continued training existing models (with some surgical changes to improve long context memory) so they should score similarly in these tests.
Comments
Note that until last generation all Chinese models were relatively small. This was mainly due to lack of training hardware, but as soon as the new Huawei NPUs started shipping some Chinese labs switched to larger models:
Deepseek: 670B to 1.6T
Moonshot: 1.1T to 2.8T
Xiaomi: 310B to 1T
The new models also use a different architecture so I would assume in tests they will look different from previous generations.
Z.ai (GLM) and MiniMax on the other hand have continued training existing models (with some surgical changes to improve long context memory) so they should score similarly in these tests.