DeepSeek updates its reasoning model as DeepSeek-R1-0528

What happened

DeepSeek released an updated version of DeepSeek-R1 on 28 May 2025 under the MIT License. ^[2][1]

Distillation means training a smaller model to imitate the outputs of a larger one. The R1 update is relevant here because its reasoning traces were used to train a smaller model, Qwen3 8B Base, as described below.

Key details

AIME 2025 accuracy rose from 70.0% to 87.5%, while the average number of tokens used per AIME question grew from about 12K to about 23K, reflecting deeper reasoning. ^[1]

GPQA-Diamond rose from 71.5 to 81.0, and LiveCodeBench (2408–2505) from 63.5 to 73.3. ^[1]

SWE-bench Verified rose from 49.2 to 57.6, and Humanity's Last Exam from 8.5 to 17.7. ^[1]

The update added system-prompt support and no longer required a forced <think> prefix to trigger reasoning. Function calling was enhanced, with a BFCL v3 multi-turn score of 37.0. ^[1]

DeepSeek distilled its chain of thought into Qwen3 8B Base as DeepSeek-R1-0528-Qwen3-8B. That model scored 86.0 on AIME 2024, 10 points above Qwen3 8B and matching Qwen3-235B-thinking. ^[1]

Impact

The MIT License for the update permits commercial use and distillation, so others can build on the model under a license that is already public. The gains are reported by DeepSeek and are best read alongside measurements made by independent parties on the same benchmarks.