Exploring Deepseek R1-0528 Open Source Model: Enhanced AI Programming Capabilities Competitive with OpenAI o3 and o4-mini

On May 29, 2025, Deepseek recently invited users to test the DeepSeek-R1-0528 model. Preliminary test results indicate that R1-0528 excels in programming capabilities, aesthetic design, and code completion. It shows high precision and efficiency, particularly in handling complex commands and generating frontend pages.

The model has achieved performance improvements in multiple areas, most notably its programming ability, where it can quickly generate high-quality code based on simple prompts from users.

Performance testing on the Live CodeBench platform demonstrates that its capabilities are comparable to OpenAI’s latest o3 model (High).

In an Extended NYT Connections benchmark test, the DeepSeek-R1-0528 scored 49.8, a significant increase from the original Deepseek R1’s score of 38.6. This benchmark evaluates large language model (LLM) performance based on the New York Times Connections puzzle game.

R1-0528 exhibits response styles akin to OpenAI o3 and Google Gemini 2.5 Pro. Its use of arrows and asterisks aligns closely with the o3 style, and its concluding statements, like ‘why it works’, are more persuasive.

Furthermore, R1-0528 has demonstrated exceptional performance in aesthetic design and code completion. It adeptly handles a wide variety of tasks, yielding accurate and practical results.

When generating complex frontend pages and dynamic animations, R1-0528 shows robust capabilities, accurately understanding complex commands. Notably, it significantly reduces the inference duration compared to OpenAI’s o3 and o4-mini models, providing a more seamless and efficient user experience.

Note: Links to external sources within this article are intended to provide additional information, and are for reference only. All articles published by IT Home include this disclaimer.