On May 28, 2025, Chinese AI startup DeepSeek released an updated version of its R1 model, named DeepSeek-R1-0528, along with a distilled variant, DeepSeek-R1-0528-Qwen3-8B. This release marks a significant step forward in the development of open-source reasoning models, offering enhanced capabilities in mathematics, programming, and logical reasoning.(The Times of India, Hugging Face)
DeepSeek-R1-0528 demonstrates notable improvements over its predecessor, particularly in handling complex reasoning tasks. For instance, in the AIME 2025 benchmark, the model’s accuracy increased from 70% to 87.5%. This advancement is attributed to deeper reasoning processes, with the model averaging 23K tokens per question, nearly doubling the previous version’s 12K tokens. Such enhancements bring DeepSeek-R1-0528 closer in performance to leading models like OpenAI’s O3 and Google’s Gemini 2.5 Pro. (Hugging Face, Wikipedia)
Building upon the advancements of DeepSeek-R1-0528, DeepSeek introduced a distilled model, DeepSeek-R1-0528-Qwen3-8B. This model leverages the chain-of-thought reasoning from R1-0528 to fine-tune the Qwen3 8B base model. The result is a compact yet powerful model that achieves state-of-the-art performance among open-source models on benchmarks like AIME 2024, surpassing Qwen3 8B by over 10% and matching the performance of larger models like Qwen3-235B-thinking. (Wikipedia, Hugging Face)
DeepSeek-R1-0528-Qwen3-8B exhibits strong performance across various benchmarks:
These results highlight the model’s proficiency in mathematical and logical reasoning tasks, positioning it as a competitive option in the open-source AI landscape. (Hugging Face)
DeepSeek-R1-0528-Qwen3-8B is available under the MIT License on Hugging Face, making it accessible for both research and commercial applications. The model supports a maximum generation length of 64K tokens and can be run locally using the same configuration as Qwen3-8B, with the added benefit of supporting system prompts without the need for special tokens to initiate reasoning patterns. (Hugging Face)
The release of DeepSeek-R1-0528 and its distilled variant underscores the potential of open-source models to rival proprietary counterparts in performance. By providing high-quality reasoning capabilities in a compact and accessible format, DeepSeek contributes to the democratization of AI technology, enabling broader participation in AI development and research.(Wikipedia, The Times of India)