DeepSeek, a Chinese AI lab, has released a new smaller version of its advanced R1 reasoning AI model called DeepSeek-R1-0528-Qwen3-8B. This “distilled” model delivers strong performance while running on much less powerful hardware, making it accessible for wider use.
Built on Alibaba’s Qwen3-8B foundation model, the distilled R1 outperforms Google’s Gemini 2.5 Flash on a tough math challenge called AIME 2025. It also nearly matches Microsoft’s Phi 4 reasoning plus model on another math test known as HMMT.
While smaller models like DeepSeek-R1-0528-Qwen3-8B usually have fewer capabilities than full-sized versions, they require far less computing power.
The full R1 model needs around a dozen GPUs with 80GB of memory, but the distilled model can run on a single GPU with 40GB to 80GB of RAM, such as an Nvidia H100.
DeepSeek created the smaller model by fine-tuning Qwen3-8B using text generated by the larger R1 model. The company describes this model as ideal for academic research and industrial use where smaller AI models are needed.
Importantly, DeepSeek-R1-0528-Qwen3-8B is available under an open MIT license, allowing commercial use without restrictions. Several platforms, including LM Studio, already provide access to the model via API.
