Model Drop

DeepSeek R1 0528

May 28, 2025 2 min read

DeepSeek R1 0528, which DeepSeek shipped on May 28, 2025, is an MIT-licensed update to its R1 reasoning model that thinks longer and moves close to o3 and Gemini 2.5 Pro on math and code.

Model: DeepSeek R1 0528 (deepseek-ai/DeepSeek-R1-0528)
Maker: DeepSeek
Ship date: May 28, 2025
Pricing: $0.55 / $2.19 per M tokens (DeepSeek API); free weights
Context: 128K tokens (64K on DeepSeek's API and app)
Weights: Open (MIT)
Where you can use it:
- DeepThink mode in DeepSeek chat
- The API as deepseek-reasoner and open weights
When to reach for it:
- Hard math, code and logic problems
- Front-end code and function calling
- Skip it when speed matters: it thinks long
How it compares:
- Close behind o3 and Gemini 2.5 Pro
- Ahead of Qwen3 235B among open models
What makes it different:
- Deeper thinking, fewer hallucinations
- MIT license; distilled 8B Qwen3 version
Early reactions: DeepLearning.AI's The Batch said the update leads all open models and "nips at the heels of top closed LLMs," though it still trails o3 on most tests.
Read more: DeepSeek-R1-0528 Release, model card

The benchmarks

DeepSeek's own runs with a 64K-token generation cap, temperature 0.6 and pass@1 averaged over 16 samples; it compared against o3, Gemini 2.5 Pro (0506 preview), Qwen3 235B and the original DeepSeek R1, with rival scores as DeepSeek printed them (no o3 effort level given). Humanity's Last Exam is the text-only subset, SWE-bench uses the Agentless framework, and the original R1's Aider score is the launch chart's 57.0 (the model card prints 53.3).

DeepSeek's numbers

AIME 2024 (competition math): 91.4%. −0.2 under o3 (91.6%), +0.6 over Gemini 2.5 Pro (90.8%), +5.7 over Qwen3 235B (85.7%), +11.6 over DeepSeek R1 (79.8%).
AIME 2025 (competition math): 87.5%. −1.4 under o3 (88.9%), +4.5 over Gemini 2.5 Pro (83.0%), +6.0 over Qwen3 235B (81.5%), +17.5 over DeepSeek R1 (70.0%).
GPQA Diamond (PhD-level science): 81.0%. −2.3 under o3 (83.3%), −2.0 under Gemini 2.5 Pro (83.0%), +9.9 over Qwen3 235B (71.1%), +9.5 over DeepSeek R1 (71.5%).
LiveCodeBench (fresh coding problems): 73.3%. −4.0 under o3 (77.3%), +1.5 over Gemini 2.5 Pro (71.8%), +6.8 over Qwen3 235B (66.5%), +9.8 over DeepSeek R1 (63.5%).
Aider Polyglot (multi-language code edits): 71.6%. −8.0 under o3 (79.6%), −5.3 under Gemini 2.5 Pro (76.9%), +6.6 over Qwen3 235B (65.0%), +14.6 over DeepSeek R1 (57.0%).
Humanity's Last Exam (no tools): 17.7%. −2.9 under o3 (20.6%), −0.7 under Gemini 2.5 Pro (18.4%), +5.9 over Qwen3 235B (11.8%), +9.2 over DeepSeek R1 (8.5%).
SWE-bench Verified (real GitHub fixes): 57.6%. +8.4 over DeepSeek R1 (49.2%).
SimpleQA (short factual recall): 27.8%. −2.3 under DeepSeek R1 (30.1%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 13.1. +1.7 over DeepSeek R1 (11.4), +3.6 over Qwen3 235B (9.5), −7.1 under o3 (20.2).