DeepSeek debuts lighter R1 AI model with better math reasoning

Just a heads up, if you buy something through our links, we may get a small share of the sale. It’s one of the ways we keep the lights on here. Click here for more.

While DeepSeek’s powerful new AI model, R1, has been grabbing headlines, the Chinese AI lab also quietly released a lighter, more efficient version of it. 

This smaller model, called DeepSeek-R1-0528-Qwen3-8B, might not be as big or powerful as the full R1, but it still performs impressively well, especially in solving complex math problems.

This new model is built on top of Qwen3-8B, which is a model released by Alibaba in May 2025. 

DeepSeek improved it by using outputs from their full R1 model to “teach” the smaller version, a technique called distillation. 

This is like training a student by giving them solutions from an expert and helping them learn the logic behind the answers.

Despite its smaller size, DeepSeek-R1-0528-Qwen3-8B beats Google’s Gemini 2.5 Flash on a tough math test called AIME 2025 and comes close to matching Microsoft’s Phi-4 reasoning model on another test, HMMT, both of which are known for challenging problem-solving tasks.

What makes this model exciting isn’t just its strong performance, but it’s much easier and cheaper to run. 

While the full R1 model requires a massive setup of multiple high-end graphics cards (GPUs), the smaller version based on Qwen3-8B can run on a single GPU with 40–80GB of RAM. (Via: Tech Crunch) 

That’s still powerful hardware, but far more manageable for smaller companies and developers.

DeepSeek is offering this model under the MIT license, which means anyone can use it for free, even in commercial products. 

The model is already available through popular platforms like LM Studio and Hugging Face, making it easy for developers to experiment with or build on top of it.

DeepSeek-R1-0528-Qwen3-8B is a smaller, faster, and more accessible AI model that still holds its own in complex reasoning tasks, especially math, making it a valuable tool for both researchers and businesses looking for efficient AI solutions.

Will you be using this new model from DeepSeek? Do you think it’ll be a good alternative to newer mainstream AI models? Tell us below in the comments, or via our Twitter or Facebook.