A May 2025 checkpoint update to DeepSeek R1 with substantially deeper chain-of-thought reasoning and large gains on math, coding, and general-knowledge benchmarks.

Benchmarks

What its maker published, and what anyone else measured. Every figure links the document it came from.

benchmark	published	measured
MMLU-Pro	85	—
GPQA	81	—
AIME 2025	87.5	—
SWE-bench	57.6	—

Lineage

↑ finetune DeepSeek R1

view in full graph →

Metadata

org	DeepSeek
released	2025-05-28
params	671B-A37B
license	mit
hf id	deepseek-ai/DeepSeek-R1-0528