Chinese artificial intelligence firm DeepSeek has announced a significant upgrade to its flagship large language model, R1, intensifying competition with industry leaders OpenAI and Google (GOOG) (GOOGL). The new version, dubbed DeepSeek-R1-0528, is now being positioned as a direct challenger to OpenAI’s o3 and Google’s Gemini 2.5 Pro, with benchmark results and technical enhancements that show China’s rapidly advancing AI sector. “Its overall performance is now approaching that of leading models, such as O3 and Gemini 2.5 Pro,” DeepSeek said in a post on HuggingFace.

Key Upgrades and Performance Gains

DeepSeek’s latest update centers on improved reasoning and inference abilities, achieved by leveraging increased computational resources and introducing advanced post-training algorithmic optimizations. The company reports that R1-0528’s accuracy on complex reasoning tasks has surged to 87.5%, up from 70% in earlier versions—a leap attributed to deeper reasoning and a substantial increase in the number of tokens processed per question.

Notably, the model’s hallucination rate — instances where the AI generates inaccurate or misleading information — has been cut by half, and its capabilities in mathematics, programming, and general logic have all seen marked improvements. DeepSeek also highlights enhanced coding support, including better function calling and a smoother experience for “vibe coding,” where developers use conversational prompts to generate code.

Open-Source Advantage and Market Impact

Unlike the closed-source models from OpenAI and Google, DeepSeek continues to release its model weights to the public, maintaining an open-source approach that has resonated with the developer community. The R1-0528 model, now boasting 685 billion parameters, is available under a permissive MIT license, allowing for commercial use and broader adoption.

Don’t Miss: 

• Think it’s too late to invest in the booming AI sector? This one’s still under the radar
• How This Startup is Disrupting the $4.83 Trillion Wealth Engine That’s Been Hidden From Everyday Investors for Over 93 Years

DeepSeek has also introduced a distilled, lighter variant of R1-0528, which can operate on significantly less computational power — potentially even a single GPU according to TechCrunch— while still outperforming Google’s Gemini-2.5-Flash-Thinking-0520 and OpenAI’s o3-mini in several benchmarks. This open-weight strategy not only democratizes access to advanced AI but also challenges the prevailing notion that only massive computational investments can yield top-tier AI performance.

Competitive Landscape: China’s AI Surge

The R1-0528 upgrade comes amid fierce competition among Chinese tech giants. Alibaba’s (BABA) Qwen 3 and Baidu’s (BIDU) Ernie 4.5 and X1 models have all been released in recent months, each vying for supremacy in the rapidly evolving AI landscape. Benchmark results cited by DeepSeek show that R1-0528 now surpasses Alibaba’s Qwen 3 and matches the performance of OpenAI and Google’s best models.

Industry observers are watching closely for the anticipated launch of DeepSeek’s R2 model, which was initially expected in May. The company’s ongoing advancements, including last month’s release of the open-sourced Prover-V2 specialist model, signal its ambition to remain at the forefront of global AI innovation.

As DeepSeek’s R1-0528 continues to close the gap with Western leaders, the global AI race is entering a new phase — one where open-source strategies, cost efficiency, and rapid technical progress are redefining the boundaries of what’s possible in artificial intelligence.