What will surprise the market with the new Chinese DeepSeek R2

      Chinese startup DeepSeek is preparing to surprise the artificial intelligence market again. This time we are talking about the R2 model, information about which has already caused a wave of discussions due to impressive technological achievements in three areas at once.

       Online

       The characteristics of the latest DeepSeek R2 model, capable, according to preliminary estimates, of bypassing industry leaders, have been leaked. DeepSeek is a leading Chinese startup in the field of AI. The company was established in 2023 with the aim of "exploring the essence of general artificial intelligence." The IT-World has studied what is expected "under the hood" of the new release.

      

       The first and perhaps the main advantage of the new development is its revolutionary Hybrid MoE 3.0 architecture. DeepSeek can manage 1.2 trillion parameters with an actual load of only 78 billion. Thanks to this optimization, the cost of data processing (tokens) has become lower by an impressive 97.3% compared to GPT-4 Turbo from OpenAI. Against the background of such figures, even the market leaders are starting to look a little outdated.

       The second key area was the achievement of high computational efficiency on domestic equipment. DeepSeek R2 showed 82% capacity utilization of the Huawei Ascend 910B chip cluster, delivering 512 PetaFLOPS of performance. This is equivalent to 91% of the power of the famous NVIDIA A100 chips, but already using Chinese technologies. It sounds intriguing and a bit defiant: can Western leaders really be left on the bench soon?

       The third breakthrough is in multimodal tasks. Here again, R2 is surprising: the accuracy of object segmentation on the well-known COCO dataset has reached 92.4%, which is almost 12 points better than the popular CLIP model. In production control, the false alarm rate dropped to an incredible 7.2E-6. And in medical diagnostics based on chest X-rays, the new model surpassed professional radiologists with an accuracy of 98.1%, exceeding their average (96.3%).

       Under the hood of DeepSeek R2 is a huge amount of 5.2 petabytes of data covering finance, law, and patents. According to the C-Eval 2.0 tests, the model shows the accuracy of instructions at 89.7%. Another advantage is the quantization technology, which reduces the size of the model by 83%, with virtually no loss of accuracy when switching to 8-bit precision. This makes R2 accessible even for devices with limited computing capabilities, expanding its application in industry, healthcare and urban management.

       The large-scale project is, of course, supported by major technology partners.: Tuowei Information provides more than half of Huawei Ascend's infrastructure, Zhongke Shuguang supplies liquid-cooled servers, Inspur Information is responsible for more than 5,000 servers with NVIDIA and Huawei hybrid chips, and Xinyisheng has developed energy-saving solutions based on silicon photonics.

       If the official data is confirmed, DeepSeek R2 has every chance of changing the balance of power in the artificial intelligence market, and very rapidly and unexpectedly for Western competitors. It seems that the AI race has just begun to enter its most interesting phase.