DeepSeek unveiled a compact version of the R1 AI model to run on a single GPU
Author: NaKmo Flow | 5/30/2025
Chinese company DeepSeek has released a new, lightweight version of its R1 language model, designated DeepSeek-R1-0528-Qwen3-8B. Unlike the original R1, the new product can efficiently run only one GPU, making it more affordable for individual developers and small teams.
The emphasis is on logic and reasoning
The new model is based on Qwen3-8B from Alibaba and is optimized for tasks requiring analysis, logical inference and reasoning. In particular, it performs well in profile tests:
- AIME 2025
- HMMT (Harvard-MIT Math Tournament)
Model Characteristics
| Parameter | Value | |
| Name | Name | DeepSeek-R1-0528-Qwen3-8B |
| Architecture | Qwen3-8B (from Alibaba) | |
| GPU Support | Single GPU | One GPU |
| Major specialization | Mathematics, logic, reasoning | |
| License | MIT (open, including commercial use) | |
| Distribution Platform | Hugging Face | |
| Language Policy | Moderately censored, with mild disclaimers |
More open but cautious
The updated version of R1 has changed its approach to moderation: previously the model could directly avoid answering politically sensitive questions (e.g. about the Tiananmen Square events or comparing leaders to pop culture characters). Now the model simply politely evades answering with the phrase:
"Sorry, that's beyond my current scope. Let's talk about something else."
This change seems to be made to balance transparency, security and accessibility in an international context.
Who is this model for?
The compact version of the DeepSeek R1 can be particularly useful:
- Researchers and students working with logic and math
- Startups that can't afford to run GPT-4 scale models
- Developers implementing AI in on-premises solutions without cloud connectivity
The model is already available for download and use at Hugging Face.