menu
close

DeepSeek Upgrades R1 AI Model, Challenges Western Tech Giants

Chinese AI startup DeepSeek has released an update to its R1 reasoning model that helped propel the company to global prominence earlier this year. The R1-0528 update, while described as 'minor' by the company, shows significant improvements in coding capabilities, reasoning depth, and writing tasks. The updated model maintains DeepSeek's cost-efficient approach to AI development while rivaling performance of models from OpenAI and Google.
DeepSeek Upgrades R1 AI Model, Challenges Western Tech Giants

DeepSeek, the Chinese artificial intelligence startup that sent tech shares tumbling in January with its groundbreaking R1 model, has quietly rolled out a significant update to its flagship AI system.

The new version, dubbed DeepSeek-R1-0528, was released on May 28 with little fanfare through a brief announcement in the company's WeChat group. Despite DeepSeek describing it as a "minor trial upgrade," early testing reveals substantial improvements across multiple domains.

According to the LiveCodeBench leaderboard, the updated model now ranks fourth in coding capabilities with a Pass@1 score of 73.1, placing it just behind OpenAI's O3 and O4-Mini models. This represents a major leap forward for an open-source model, especially one developed with significantly fewer resources than its Western counterparts.

Technical specifications show R1-0528 maintains the Mixture-of-Experts (MoE) architecture of its predecessor, with approximately 685 billion parameters in total, though only about 37 billion are active during inference. The model now supports an expanded context window of up to 128K tokens, enabling it to process much larger documents and codebases.

User feedback highlights improvements in reasoning depth, writing quality, and problem-solving capabilities. Developers report the model can now engage in extended reasoning sessions lasting 30-60 minutes for complex tasks, similar to Google's models. The update also addresses previous quirks in text generation, producing more natural and better-formatted content.

DeepSeek's continued innovation challenges the notion that scaling AI requires vast computing power and investment. The company's success with cost-efficient, open-source models has already forced responses from tech giants, with OpenAI cutting prices and Google introducing discounted tiers of access. Meanwhile, Chinese competitors like Alibaba and Tencent have released their own models claiming to surpass DeepSeek's capabilities.

The R1-0528 update is available on Hugging Face under an MIT license, allowing for commercial use and modifications. Industry watchers anticipate DeepSeek will release its more comprehensive R2 model in the coming months, potentially further disrupting the AI landscape.

Source:

Latest News