The Chinese startup DeepSeek has introduced an updated version of its R1 language model on the Hugging Face platform, causing a real stir in the AI community. The update, named R1-0528, is notable for its impressive size of six hundred eighty-five billion parameters, yet the model is available under an open MIT license, allowing for free commercial use.
DeepSeek also released a more compact “distilled” version of the model — DeepSeek-R1-0528-Qwen3-8B, which is based on Qwen3-8B from Alibaba. Despite its smaller size, this model demonstrates brilliant results: it surpasses Google Gemini 2.5 Flash in complex mathematical tasks AIME 2025 and nearly matches Microsoft Phi 4 reasoning plus in tests of other skills. Meanwhile, DeepSeek-R1-0528-Qwen3-8B is significantly less demanding on computational resources, opening up new opportunities for developers and companies.
However, independent tests have shown that R1-0528 has become noticeably stricter on content related to political topics considered sensitive by Chinese authorities. It is known that the model often avoids answering questions about Xinjiang and, in some cases, repeats the official government stance, even when it comes to human rights violations. Developers note that this is the most censored version of DeepSeek ever released.
DeepSeek has previously been in the spotlight due to strict limitations in its models. The company emphasizes the scientific and industrial potential of new solutions, but the question of the independence of their models sparks active discussion in the global AI developer community.
