ChatGPT-5 has been released, so a major version update of the domestic large model DeepSeek is not far away.Judging from the questions asked by netizens to DeepSeek, the answer given by the model is that DeepSeek-R2 is expected to be released between August 15 and August 30, 2025.
Against this background, the stock prices of many companies in the domestic computing power chain have surged. For example, Cambrian has reached a daily limit of 20cm today, with its stock price hitting a record high and its market value exceeding 355 billion yuan.
According to previous statements, the DeepSeek-R2 large model will use a more advanced hybrid expert model (MoE), which is combined with a more intelligent gating network layer (Gating Network) to optimize the performance of high-load inference tasks.
Some analysts predict that the pricing of DeepSeek-R2 may be significantly lower than similar products of OpenAI, indicating that it may subvert the pricing model of existing AI services.
In addition, relevant news shows that DeepSeek-R2 is expected to cost 97% less than GPT-4; and it is trained on the Ascend card, focusing on a comprehensive and independent control of the entire industry chain.
According to relevant sources, the total number of parameters of DeepSeek-R2 may reach 1.2 trillion, which is approximately double the 671 billion parameters of DeepSeek-R1.
DeepSeek-R2 is a model trained based on Huawei's Ascend 910B chip cluster. It has a computing power of 512PetaFLOPS at FP16 accuracy and a chip usage efficiency of 82%.
According to statistics from Huawei Labs, this performance is equivalent to about 91% of NVIDIA’s previous generation A100 training cluster. Whether or not these are real remains to be seen, but it does appear to be more powerful and autonomous.