Deepseek-R1-0528 Update Official Detailed Explanation: Deeper Thinking, Stronger Reasoning

The Zhitong Finance App learned that this evening, DeepSeek officially announced the details of the DeepSeek-R1-0528 update. Deepseek-R1-0528 still uses the DeepSeek V3 Base model released in December 2024 as a base, but more computing power was invested in the post-training process, which significantly improved the model's depth of thought and reasoning ability. The updated R1 model achieved superior results among all current domestic models in multiple benchmark evaluations such as mathematics, programming, and general logic, and is close to other top international models in terms of overall performance, such as O3 and Gemini-2.5-Pro.

Deepseek-R1-0528 achieved excellent results in all evaluation sets. Compared to the previous version of R1, the performance of the new model has improved significantly in complex inference tasks. For example, in AIME 2025 testing, the accuracy of the new model increased from 70% of the previous version to 87.5%.

This progress is due to the model's increased depth of thought in the reasoning process: on the AIME 2025 test set, the old model used an average of 12K tokens per question, while the new model used an average of 23K tokens per question, indicating that they thought more thoroughly and deeply during the problem solving process.

Additionally, the new DeepSeek R1 is optimized for “illusion” issues. Compared with the previous version, the updated model reduces the illusion rate by 45 to 50% in scenarios such as rewriting and refinement, summarizing, and reading comprehension, and can effectively provide more accurate and reliable results. Based on the old version of R1, the updated R1 model has been further optimized for styles such as argumentative essays, novels, and essays. It can output long-form works with longer length and more complete structural content, and at the same time present a writing style closer to human preferences.