Small model Deepseek R1 has updated and already hit Gemini 2.5 flash in the benchmark test – computers
Chinese startup Deepsek continues to improve its artificial intelligence model, which is the distillation R1 version, which has received a new update. The version is known as Deepsek-R1-0528-QWEN3-8B and is shown in small and more brief. However, the benchmark AIME 2025 test results show that this is Google’s latest model, Gemini is very effective and thin against Opena 03 in 2.5 flash performance.
Perhaps the most surprising thing is that this version of Deepseek R1 requires less hardware resources to work. According to Tech CrunchThis version of R1 uses Alibaba’s QWEN3-8B and requires NVIDIA H100 to operate. Compared, as per the cloud platform Node ShiftA full version requires a dozen 80 GPUs to work for Deepsek R1.
Deepseek trained the new model using the updated R1 version using the text. The There is no startup Hugged the face This update significantly improves its logical and suspicion capabilitiesBy utilizing the increase in computation resources and introducing the procedures of the post-workout algorithm’s optimization.
Oh New model exhibits high performance between various benchmark assessments, including mathematics, programming and general logic. “Your normal performance will now reach prominent models like O3 and Gemini 2.5 Pro,” he pointed out in his publication. DEPSECK-R1-0528-QWEN3-8B says that it can be used in academic research and industrial development focused on small-scale models.
Note that the model is available by A MIT LicenseCan be used in commercial products without limitsOnly refer to the refusal of use.
