Atlas

Models

← All models

DeepSeek-R1-0528

DeepSeek · 2025-05-28 · 684.5B parameters

DeepSeek-R1-0528 is the May 28, 2025 minor-version update to DeepSeek-R1, retaining its DeepSeek-V3 MoE backbone (reported as 671B main-model parameters with 37B activated per token; the released checkpoint also contains one multi-token-prediction module). DeepSeek attributes stronger reasoning to increased post-training compute and algorithmic optimization; on AIME, average reasoning length rose from 12K to 23K tokens. The release also reduces hallucinations and improves system-prompt, JSON-output, function-calling, and coding support.

Benchmark scores

BenchmarkScore
AA-LCR54.7
AA-LCR v1.155.7
AA-Omniscience Index-27.4
AIME 202587.5
AIME 2025 (Artificial Analysis avg@10)76.0
AIME 2025 (MathArena repeated-run mean)89.2
ALE-Bench1059
AlgoTune1.7
Apex (MathArena)1.0
ARC-AGI-1 Semi-Private21.2
ARC-AGI-21.1
Arena Text — Business, Management & Financial Ops (No Style Control)1401
Arena Text — Chinese (No Style Control)1442
Arena Text — Creative Writing (No Style Control)1405
Arena Text — English (No Style Control)1437
Arena Text — Entertainment, Sports & Media (No Style Control)1400
Arena Text — Exclude Ties (No Style Control)1420
Arena Text — Exclude Ties (Style Controlled)1412
Arena Text — Expert (No Style Control)1393
Arena Text — French (No Style Control)1417
Arena Text — French (Style Controlled)1420
Arena Text — German (No Style Control)1404
Arena Text — Hard Prompts (No Style Control)1415
Arena Text — Hard Prompts English (No Style Control)1422
Arena Text — Instruction Following (No Style Control)1381
Arena Text — Japanese (No Style Control)1391
Arena Text — Korean (No Style Control)1359
Arena Text — Legal & Government (No Style Control)1419
Arena Text — Life, Physical & Social Science (No Style Control)1430
Arena Text — Longer Query (No Style Control)1391
Loading Atlas data…