DeepSeek R1 (0528)

DeepSeek announced the DeepSeek-R1-0528 release on May 28, 2025, as a minor but meaningful upgrade to its R1 reasoning line. The official post lists improved benchmark performance, enhanced front-end capabilities, and a reduced hallucination rate as headline gains. DeepSeek also added support for JSON output and function calling to the model. API usage remains unchanged from prior R1 versions, with a Thinking Mode guide linked for developers. Open-source weights were published to Hugging Face under deepseek-ai/DeepSeek-R1-0528. The chat interface at chat.deepseek.com was updated to serve the new variant immediately.

The DeepSeek API Docs published a release note for DeepSeek-R1-0528 on May 28, 2025, confirming the rollout to developers. The note highlights improved benchmark performance, enhanced front-end capabilities, reduced hallucinations, and new support for JSON output and function calling. The post links to the chat.deepseek.com interface and reiterates that API usage is unchanged, pointing to the Thinking Mode guide for details. Open-source weights are linked from Hugging Face, mirroring the official DeepSeek news announcement for the same variant.

The Unsloth mirror of the DeepSeek-R1-0528 model card documents the upgrade's reasoning gains. The card reports the new model approaches the performance of leading systems like O3 and Gemini 2.5 Pro, with AIME 2025 accuracy rising from 70% to 87.5% and average reasoning tokens roughly doubling to 23K per question. Additional capability gains noted include a reduced hallucination rate, improved function calling support, and a better vibe coding experience. Detailed benchmark tables show gains across MMLU-Pro, GPQA-Diamond, Humanity's Last Exam, LiveCodeBench, Codeforces, SWE Verified, Aider, AIME, HMMT, and CNMO versus the prior R1 release.