DeepSeek has introduced an upgraded AI model, the DeepSeek-R1-0528, which shows promise in outperforming competitors like ChatGPT and Gemini based on initial benchmark tests. This model outperforms Gemini’s free version and nearly matches the performance of OpenAI’s o3 model. Additionally, DeepSeek has released a “distilled” variant called DeepSeek-R1-0528-Qwen3-8B, which is designed for efficiency, requiring significantly fewer resources while maintaining robust performance.
The lightweight version, based on Alibaba’s Qwen3-8B model, has excelled in the AIME 2025 test by solving complex mathematical problems faster than Google’s Gemini 2.5 Flash and has matched Microsoft’s Phi 4 reasoning model in the HMMT evaluations. Notably, while the full DeepSeek-R1-0528 requires over 12 GPUs with 80GB RAM each, the distilled version can operate on a single GPU with just 40GB of RAM, demonstrating its resource efficiency.
This balance of high performance with low resource demands makes the DeepSeek-R1-0528-Qwen3-8B model appealing for both academic research and small-scale industrial applications. Further enhancing its attractiveness, this model is released under a permissive MIT license, allowing unrestricted commercial use, distinguishing it from the proprietary limitations of OpenAI and Google’s models.
Source link
