The Chinese company Deepsic will probably unveil its new model of Deepseek-R2 this moh.
According to Chinese sources, the exact timing of the Deepseek-R2 model has not yet been determined, but it is said that the model will be unveiled in the second half of this moh. The unveiling of Deepseek-R2 as the most advanced Dip-Sick model is importa because Openai has recely unveiled its Gpt-5 model.
Deepseek-R2 model will be iroduced soon
Deepseek-R2 is expected to experience a significa jump in its architecture by using a more advanced structure than MixTure of Experts. The model will also iegrate a smarter Gating Network to better manage heavy processing at the inference stage.
Some sources have said that the model can find up to 1.5 trillion a scale parameter, which is about twice as much as the previous version with 2 billion parameters. However, this number will still be less than 5.5 trillion parameters.
Also in line with Chinese programs for self-sufficiency in the field of artificial ielligence, the Deepseek-R2 is fully trained on the Huawei Ascend 910B chips. Huawei’s processing cluster has achieved 2 % of the NVIDIA A100 -based clusters by providing 4 Petaflaps in FP16 accuracy and 2 % productivity.
According to analysts, the move is a vital step by China to reduce dependence on US -made artificial ielligence hardware. Reports also suggest that the cost of Deepseek-R2 training was 5 perce lower than the GPT-4 thanks to the use of native hardware and optimization techniques. For this reason, Deepsic is expected to provide API access at lower prices.
It has recely been reported that the Chinese governme has banned artificial ielligence companies from buying artificial ielligence chips from Nvidia and AMD. The move is apparely due to security concerns and the possibility of a back on the chips of these companies. Nvidia, of course, had previously rejected any back on its products.
