DeepSeek-AImodel systemcard
DeepSeek V4 Flash
#20of 27 · general
Standing
dot = this model · shaded = the field's middle 60% · track = full fieldGeneral
20.4#20/27
Math
6.8#27/27
Coding
16.0#25/27
Health
21.0#21/26
Engineering
25.0#13/17
Cyber
30.8#10/14
Legal
25.2#22/26
Instruct
20.3#22/27
Agents
33.6#15/27
Professional
37.9#18/27
Hallucination
12.6#23/27
Speed
66.7#5/22
Value
40.3#19/27
Open
44.2#3/11
Measured
≈ imputed · amber dot = lab-self-reported among the sourcesGPQA Diamond90.8%#14
HLE41.8%#14
MMLU-Pro86.2%#19
LiveBench65.5%#20
SWE-bench V79.0%#11
LiveCodeBench91.6%#1
SciCode49.9%#19
CritPt16.6%#11
Terminal-Bench67.8%#16
GDPval52.9%#10
Arena1438#20
AA Index51.8#12
Epoch Capabilities Index152.5#15
Operations
$0.14 / $0.28$ per 1M tokens, in / out
$0.18blended
124tokens / sec
$0.02$ per solved task
1438arena elo · 46,445 votes
Series
DeepSeek's models on the board, by releaseDeepSeek V4 Pro2026-04-23 · 23DeepSeek V4 Flash2026-04-23 · 20