Chinese artificial intelligence (AI) firm DeepSeek has unveiled its latest generative AI model, DeepSeek-R1-0528.
This new iteration represents a minor upgrade to the existing DeepSeek R1, boasting an impressive 685 billion parameters that significantly enhance its reasoning capabilities. The model showcases improved code generation and execution abilities, with the capacity for sustained, in-depth reasoning on specific tasks for up to 30 to 60 minutes, according to a report by GigaGen on Thursday.
DeepSeek asserts that its new model performs deep reasoning comparable to that of Google AI, offering swift yet comprehensive analyses. In LiveCodeBench, measuring code generation, modification, and output prediction, DeepSeek-R1-0528 secured the fourth position, demonstrating performance on par with OpenAI’s o4-mini. To validate the model’s reasoning abilities, DeepSeek is conducting various experiments, including tasking it with summarizing the research paper “Attention Is All You Need,” which outlines the Transformer architecture.
The DeepSeek-R1-0528 model will be freely accessible under the MIT license, enabling anyone to download and utilize the model data. As DeepSeek positions itself as a direct competitor to OpenAI in the rapidly evolving AI market, industry observers are keenly watching to see how this new model will impact the landscape.