DeepSeek Officially Announces R1 Upgrade: Improved Depth of Thought and Reasoning, Reduced Hallucination Rate by Nearly 50%

Domestic large modeling companies seek depth (DeepSeek) issue detailed upgrade announcements explainingDeepSeek-R1The specifics of the upgrade.

On May 29, according to a DeepSeek announcement, theDeepSeek R1The model has been upgraded to DeepSeek-R1-0528, users can experience the latest version by turning on the "Deep Thinking" function after entering the dialog interface through the official website, APP or apps, and the API has been synchronized and updated, and the calling method remains unchanged.

The announcement states that DeepSeek-R1-0528 uses the DeepSeek V3 Base model released in December 2024 as its base, but invests more computing power in the post-training process to significantly improve the model's depth of thinking and reasoning ability. The updated R1 model has achieved excellent results in several benchmarks in mathematics, programming and general logic, and is close to other top international models such as o3 and Gemini-2.5-Pro in terms of overall performance.

Compared to the old R1 version, the new version of the model shows significant performance improvement in complex reasoning tasks. For example, in the AIME2025 test, the accuracy of the new model increases from 701 TP4T in the old version to 87.51 TP4T.This improvement is attributed to the enhanced depth of thinking of the model during the reasoning process: in the AIME2025 test set, the old model uses an average of 12K tokens per question, while the new model uses an average of 23K tokens per question, which suggests that it has conducted more detailed and in-depth thinking during the solving of the questions. more detailed and deeper thinking in the process of problem solving.

DeepSeek said.DeepSeek-R1-0528's chain of thought will be important for both research in inference modeling in academia and development for small models in industry.

The new version of DeepSeek R1 has been optimized for the "illusion" problem. Compared with the old version, the updated model reduces the illusion rate by about 45-50% in the scenarios of rewriting, summarizing, reading comprehension, etc., which can effectively provide more accurate and reliable results.

Based on the old version of R1, the updated R1 model is further optimized for the genres of argumentative essays, novels, and prose, and is capable of outputting longer works with more complete structural content, as well as presenting a style of writing that is more closely aligned with human preferences.

According to the announcement, the current model's measured performance is comparable to OpenAI o1-high, but still falls short of o3-High and Claude4 Sonnet.

After this R1 update, the model context length in the official website, applets, apps and APIs is still 64 K. If users have a need for a longer context length, they can call the open source version of R1-0528 model with a context length of 128 K through other third-party platforms. Consistent with the older version of DeepSeek-R1, this upgrade remains an open source model, allowing users to train other models using model outputs, through model distillation, etc.

Previously on May 28, DeepSeek released a message in the official WeChat group that DeepSeekR1 model has completed the "small version of the trial upgrade", welcome to the official web page, APP, small program test (open the depth of thinking), the API interface and the way of use remains unchanged.

According to netizen reviews, the upgrade is amazing. From the feedback of social media, netizens are most concerned about the longer thinking time after this DeepSeekR1 update. According to the evaluation, this DeepSeek single-task processing time can be up to 30-60 minutes.