Chinese startup DeepSeek announced a distilled version of its advanced artificial intelligence model R1 with reasoning capabilities, which can operate on a single graphics card instead of the dozen GPUs required for the full version.
G. Ostrov
Chinese company DeepSeek has made a significant breakthrough in artificial intelligence by introducing a compact version of its reasoning-capable R1 model. The new model DeepSeek-R1-0528-Qwen3-8B was created using distillation methods and requires significantly fewer computational resources compared to the full-scale version.
Technical Specifications and Advantages
The distilled version is based on Alibaba's Qwen3-8B model, released in May 2025. Despite its compact size, the new model demonstrates impressive testing results, outperforming Google's Gemini 2.5 Flash in the AIME 2025 mathematical benchmark and nearly matching Microsoft's Phi 4 Plus in the HMMT test.
Hardware Requirements
One of the main advantages of DeepSeek-R1-0528-Qwen3-8B is its minimal hardware requirements. According to cloud platform NodeShift data, the model can operate with a single GPU having 40-80 GB of RAM, such as the Nvidia H100. In comparison, the full R1 version requires approximately twelve GPUs with 80 GB of memory each.
Training Process and Licensing
In creating the compact model, developers used text generated by the full-scale R1 version for fine-tuning the base Qwen3-8B model. The model is distributed under the permissive MIT license, allowing free commercial use.
Practical Applications
DeepSeek-R1-0528-Qwen3-8B is suitable for both academic research and industrial development focused on small-scale models. Several applications, including LM Studio, have already integrated this model through APIs, making it accessible to a broader range of developers.
Official DeepSeek website: https://www.deepseek.com/
If you encounter any problems, contact us, we will help quickly and efficiently!