The new model and style of Deepseek runs with a graphics card

The Chinese Artificial Intelligence Laboratory Deepseek has introduced a lightweight version of its R1 argument that can run with only one GPU. Released under the name Deepseek-R1-0528-QWen3-8B, the model is based on Alibaba’s QWen3-8B model, and in some mathematical tests, it has performed better than coincidental models such as Google Gemini 2.5.

This lightweight model, used by the full version of the trained R1, is much less costly and requires only 1 GB of graphics card, while the original R1 version requires about 2 H100 cards. Deepseek has released this model with MIT license, making it possible to use it without restriction.