DeepSeek-AImodel systemcard

DeepSeek V4 Pro

#19of 27 · general

thinking mode collapsed

1.6Tparams · MoE

1Mcontext

32Ttraining tokens

MITlicense

2026-04-23released

weights ↗repo ↗

Standing

dot = this model · shaded = the field's middle 60% · track = full field

General

23.0#19/27

Math

31.9#19/27

Coding

21.7#22/27

Health

6.2#24/26

Cyber

38.5#9/14

Legal

12.4#24/26

Instruct

38.9#14/27

Agents

31.6#16/27

Professional

38.3#17/27

Hallucination

25.1#15/27

Speed

59.5#7/22

Value

39.3#20/27

Open

42.3#4/11

Measured

≈ imputed · amber dot = lab-self-reported among the sources

GPQA Diamond90.3%#18

HLE42.9%#11

MMLU-Pro87.4%#13

LiveBench72.6%#16

SWE-bench V77.6%#17

LiveCodeBench90.5%#2

SciCode50.0%#18

CritPt12.9%#15

Terminal-Bench64.8%#18

APEX Agents24.3%#15

GDPval40.3%#15

Arena1457#17

AA Index45.3#17

Vals Index55.6#16

Epoch Capabilities Index149.1#19

Operations

$0.43 / $0.87$ per 1M tokens, in / out

$0.54blended

71tokens / sec

$0.05$ per solved task

1457arena elo · 46,848 votes

Series

DeepSeek's models on the board, by release

DeepSeek V4 Pro2026-04-23 · 23DeepSeek V4 Flash2026-04-23 · 20

Neighbors

nearest capability fingerprints · deltas describe the neighbor

Inkling→ weaker professional · pricier HY3→ weaker math · cheaper MiniMax M3→ near-twin DeepSeek V4 Flash→ weaker math · cheaper Gemini 3.5 Flash→ follows instructions better · pricier