DeepSeek Quietly Updates R1 Model, Tweaks Chain-of-Thought Behavior

Chinese AI startup DeepSeek has released a "minor update" to its flagship R1 model, introducing subtle changes to its reasoning behavior.

DeepSeek has not officially detailed the modifications, but early testers have observed alterations in the model's chain-of-thought (CoT) outputs.

The update was first noted by AIBattle on X, stating:

"Deepseek has released a minor update to their R1 model, now live. Notably, the Chain-of-Thought (CoT) behavior appears to have changed."

Screenshot of AIBattle's post on X

This adjustment follows DeepSeek's initial release of the R1 model in January 2025, which garnered attention for its performance in tasks involving mathematics, coding, and reasoning.

The R1 model, utilizing a Mixture of Experts (MoE) architecture, was notable for its efficiency and cost-effectiveness compared to Western counterparts.

Reportedly, the recent update appears to be a trial upgrade, with the company inviting users to test the revised model.

While specific details remain scarce, the AI community is closely monitoring the changes to assess their impact on performance and reasoning capabilities.

DeepSeek has not officially commented on the update.