While most eyes are fixed on DeepSeek’s high-powered R1 reasoning model, the Chinese AI lab has quietly dropped a compelling companion, DeepSeek-R1-0528-Qwen3-8B.
DeepSeek-R1-0528-Qwen3-8B is built on Qwen3-8B, a foundation model released by Alibaba in May 2025. By fine-tuning it with text outputs generated by the more powerful updated R1 model, DeepSeek managed to create a compact reasoning AI that beats some of its more resource-heavy peers in benchmark tests.
According to DeepSeek, their new distilled model “outperforms Google’s Gemini 2.5 Flash on AIME 2025” a benchmark featuring difficult math questions and nearly matches Microsoft’s Phi-4 Reasoning Plus on the HMMT test, another math-focused evaluation.
Distilled models are designed to retain the strengths of larger models while cutting down on the hardware requirements needed for inference. While they’re often seen as compromises, DeepSeek’s approach is showing that efficiency doesn’t have to come at the expense of performance.
For instance, Qwen3-8B can run on a single Nvidia H100 GPU with 40GB–80GB of memory, while the full-sized DeepSeek R1 model demands around twelve 80GB GPUs a stark contrast in cost and accessibility.
This makes DeepSeek-R1-0528-Qwen3-8B ideally suited for academic research and small-scale industrial applications, especially in use cases where robust math reasoning is required without access to massive compute infrastructure.
Unlike many state-of-the-art AI models encumbered by restrictive licenses, DeepSeek-R1-0528-Qwen3-8B is openly available under the permissive MIT license. That means developers and companies can freely use, adapt, and integrate it into commercial products with no usage restrictions.
As DeepSeek puts it, the model is “for both academic research on reasoning models and industrial development focused on small-scale models.”
The model is already being hosted by several services, including LM Studio, and is available via Hugging Face, giving it an accessible launchpad into real-world applications.
