The AI community is currently abuzz with DeepSeek’s newly updated R1 reasoning AI model. However, it’s worth noting that the Chinese AI lab has also introduced a more compact, “distilled” version of this model, known as DeepSeek-R1-0528-Qwen3-8B, which reportedly outperforms similarly sized models in certain benchmarks, according to DeepSeek.

This updated, smaller R1 model was developed using Alibaba’s Qwen3-8B model, launched in May, as its foundation. It demonstrates superior performance compared to Google’s Gemini 2.5 Flash on the AIME 2025 benchmark, which consists of a set of challenging math questions.

Additionally, DeepSeek-R1-0528-Qwen3-8B closely matches the performance of Microsoft’s recently released Phi 4 reasoning plus model on the HMMT math skills test.

Distilled models, like DeepSeek-R1-0528-Qwen3-8B, typically have reduced capabilities compared to their full-sized counterparts but require significantly less computational power. As noted by NodeShift, a cloud platform, Qwen3-8B can run on a GPU with 40GB-80GB of RAM, such as an Nvidia H100, whereas the full-sized new R1 model necessitates around a dozen 80GB GPUs.

To train DeepSeek-R1-0528-Qwen3-8B, DeepSeek utilized text generated by the updated R1 model to fine-tune Qwen3-8B. The model is described on the Hugging Face AI dev platform as suitable for both academic research on reasoning models and industrial development focused on small-scale models.

DeepSeek-R1-0528-Qwen3-8B is available under the permissive MIT license, allowing for unrestricted commercial use. The model can be accessed through an API provided by several hosts, including LM Studio.