What's New in DeepSeek R1-0528? 207

小编 · 2025-05-29

On May 28, DeepSeek R1 received a minor upgrade to DeepSeek-R1-0528. Benchmark results on Live CodeBench show its performance rivals OpenAI's latest high-end o3 version. Users can access it via the official website, APP, or mini-program by enabling the "Deep Thinking" mode. The API has also been updated synchronously with no change in the calling method. The key updates include:

What's New in DeepSeek R1-0528?

On May 28, the DeepSeek R1 model completed a minor version upgrade to DeepSeek - R1 - 0528. As shown in Live CodeBench, its performance can rival the latest high-version o3 from Openai. You can experience it through the official website, app, or mini-program by entering the chat interface and enabling the "Deep Thinking" feature. The API has also been updated simultaneously, with no changes to the calling method. The main aspects of this feature update are as follows:

DeepSeek R1-0528.webp

Enhanced Thinking Capability

Enhanced Thinking Depth: Based on the DeepSeek V3 Base model, with more computing power invested in optimization, the new version has achieved excellent results in tests on mathematics, coding, and general logic, approaching the international top level. For example, in the AIME 2025 test, the accuracy rate increased from 70% to 87.5%, and the average number of tokens used per question also increased significantly.
Chain-of-Thought Distillation: After distilling the chain-of-thought from DeepSeek-R1-0528 to train the Qwen3-8B Base, the resulting 8B model performs second only to DeepSeek-R1-0528 in the math test AIME 2024, surpassing Qwen3-8B and comparable to Qwen3-235B.

DeepSeek R1-0528.webp

Hallucination Optimization

Reduced Hallucination Rate: In scenarios such as rewriting and polishing, summary, and reading comprehension, the hallucination rate has been reduced by 45% to 50%, resulting in more accurate and reliable outputs.

Creative Writing Optimization

Style Optimization: Styles including argumentative essays, novels, and prose have been optimized, enabling the output of longer, more structurally complete long-form works with a style closer to human preferences.

Tool Calling Support

Tool Calling Support: Tool calling functionality is now supported, performing well in the Tau-Bench evaluation, comparable to OpenAI o1-high, but does not support tool calling during thinking.

API Updates

Unchanged Interface and Calling Method: The API has been updated while keeping the interface and calling method unchanged, but with added support for Function Calling and JsonOutput.
Parameter Adjustment: The meaning of the max_tokens parameter has been adjusted to limit the total length of the model's single output, including the thinking process.

Other Capability Enhancements

Front-End Code Generation: Capabilities in front-end code generation, role-playing, and other areas have been improved.
Interactive Animation Generation: Newly added support for interactive animation generation and demonstration features, such as dynamic visualization of activation parameters.
Dynamic Content Updates: Faster response and more refined details in dynamic-update front-end tasks (such as real-time weather information cards).