DeepSeek has released a smaller version of its R1 model that requires only one GPU to operate.
The new distilled version of the R1 model, called DeepSeek-R1-0528-Qwen3-8B, according to the Chinese AI company, outperforms comparable models on some benchmarks. It is built using Alibaba’s Qwen3-8B model as a base.
Less Computing Power
DeepSeek claims that this model outperforms Gemini 2.5 and Phi-4-Reasoning on mathematical benchmarks. Smaller models are often less capable than the ‘normal’ versions, but they also require significantly less computing power. According to cloud platform NodeShift, this model only needs one GPU with between 40 GB and 80 GB of RAM to operate. In comparison, the normal R1 model requires about 80 GPUs.
The model is trained with generated text from the recently updated R1 model. DeepSeek states that this model has a lower hallucination rate, extensive support for function calls, and a better coding experience.
Not everyone is a fan of the Chinese AI technology. Both the Belgian and American governments prohibit their personnel from using DeepSeek, and the app would even be banned from Google and Apple’s app stores in the US. The company is also said to be lax with security and privacy.
read also