Archive — US — techcrunch
DeepSeek’s distilled new R1 AI model can run on a single GPU
DeepSeek has introduced a new distilled AI model, DeepSeek-R1-0528-Qwen3-8B, which outperforms similarly sized models on specific benchmarks. This model requires significantly less computational power than its full-sized counterpart, making it accessible for industrial and academic research. It is available under an MIT license for commercial use.
Key points
- DeepSeek released the smaller AI model DeepSeek-R1-0528-Qwen3-8B, based on Alibaba's Qwen3-8B.
- The distilled model surpasses benchmarks of models like Google’s Gemini2.5 Flash and nearly matches Microsoft’s Phi4.
- Distilled models are usually less capable than full-size versions, but are less demanding computationally.
- DeepSeek-R1-0528-Qwen3-8B is trained using text generated from the updated R1 model.
- The model is accessible for both industrial and academic use and is available via an API under a permissive MIT license.
Why it matters
- The shift towards more efficient AI models aligns with growing trends in AI to reduce computational costs and environmental impacts.
- The development reflects global interest in making AI accessible for various users, particularly in light of increasing computational requirements of larger models.
- By 2025, the IT landscape is expected to emphasize smaller yet effective AI models, enhancing performance without the need for extensive resources.
DeepSeek-R1-0528-Qwen3-8B represents an important step towards creating more efficient AI solutions, making advanced reasoning capabilities accessible while minimizing resource requirements, which is crucial in today's rapidly evolving tech landscape.