DeepSeek-V4-Flash

DeepSeek-V4-Flash places in the top three on 7 benchmarks.

DeepSeek Released April 24th, 2026 Compare with others

121

Benchmarks with a score

Ranked first, out of five or more models

In the top three

DeepSeek Apr 2026

DeepSeek-V4-Flash

Open weights

146.8 #86 of 256

Epoch Capabilities Index

Self-Reported Index 150.0 #93 of 404

DeepSWE 53.3% #20 /28
AutomationBench 18.1% #18 /42
Toolathlon 48.2% #10 /50

Where it ranks

Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.

Knowledge and language

Benchmark	Score	Rank	Source
MMLU-Pro	86.2%	#17 of 167	Lab model card
HUMAINE	32	#7 of 50	Official leaderboard
Chinese SimpleQA	78.9%	#3 of 13	Lab model card
LMArena (Elo)	1,439	#79 of 335	Official leaderboard
Extended NYT Connections	89.6%	#25 of 103	Official leaderboard
EQ-Bench 3	1,504	#22 of 78	Official leaderboard
CulturalRiddles	39.2%	#6 of 20	Official leaderboard
Wer wird Millionär? LLM Benchmark	61867.0	#24 of 77	Official leaderboard
MRCR v2	78.7%	#10 of 30	Lab model card
StudyArena Text	1,242	#31 of 85	Official leaderboard
Longform Creative Writing (EQ-Bench)	66.0	#53 of 119	Official leaderboard
HLE Extended	36.1%	#14 of 31	Official leaderboard
SEA-HELM	69.8	#33 of 71	Official leaderboard
Humanity's Last Exam	34.8%	#73 of 154	Lab model card
Creative Writing v3	69.8	#57 of 120	Official leaderboard
AA-LCR Extended	61.8%	#26 of 49	Official leaderboard
Phare	63.7%	#40 of 62	Official leaderboard
OpenAI MRCR v2	31.3%	#33 of 47	Official leaderboard
SimpleQA Verified	34.1%	#16 of 22	Lab model card
BullshitBench	16.4%	#80 of 107	Official leaderboard
BullshitBench V2	18%	#99 of 129	Official leaderboard
Judgemark v4	36.8	#31 of 40	Official leaderboard
DesignArena ASCII Art	1,104	#64 of 77	Official leaderboard
CorpusQA 1M	60.5%	#2 of 2	Lab model card

Math and science

Benchmark	Score	Rank	Source
MathArena Apex Shortlist	89.4%	#5 of 38	Official leaderboard
GPQA Diamond	88.1%	#52 of 302	Lab model card
IMO-AnswerBench	88.4%	#8 of 37	Lab model card
MathArena (Apex, IMO, USAMO)	27.1%	#11 of 47	Official leaderboard
HMMT 2026	93.9%	#11 of 31	Official leaderboard
SimpleBench	61.1%	#39 of 106	Official leaderboard
DTBench	86.4%	#68 of 180	Official leaderboard
PutnamBench	81%	#3 of 7	Lab model card
xbench-ScienceQA	71%	#14 of 29	Official leaderboard
AIME 2026	95.8%	#17 of 31	Official leaderboard
Verifying the Verifiers (Project APE CRED benchmark)	80.5%	#13 of 23	Official leaderboard
ASI-Bench	31.9	#8 of 11	Official leaderboard
EurekaBench	14.6	#6 of 7	Official leaderboard

Coding

Benchmark	Score	Rank	Source
LiveCodeBench v6	91.6%	#5 of 90	Lab model card
Codeforces rating	3,052	#4 of 47	Lab model card
StructEval	78.3%	#4 of 28	Official leaderboard
SWE-bench Verified	79%	#23 of 155	Lab model card
Code Arena	1,582	#38 of 128	Official leaderboard
APEX-SWE	46.7%	#23 of 65	Official leaderboard
SWE Atlas Codebase QnA Extended	18%	#15 of 38	Official leaderboard
SWE-bench Multilingual	73.3%	#36 of 73	Lab model card
DesignArena Game Dev	1,187	#71 of 143	Official leaderboard
DesignArena Website	1,200	#76 of 151	Official leaderboard
DesignArena 3D Design	1,171	#70 of 136	Official leaderboard
DesignArena Code	1,196	#75 of 143	Official leaderboard
WeirdML	45.6%	#60 of 112	Official leaderboard
EvoCode-Bench v2	12.2%	#7 of 13	Official leaderboard
ALE-Bench	1,380	#68 of 126	Official leaderboard
Claw-SWE-Bench	70.3%	#9 of 16	Official leaderboard
DesignArena SVG	1,146	#58 of 103	Official leaderboard
WebDev Arena	1,431	#62 of 105	arena.ai, via Epoch AI
DesignArena UI Component	1,155	#87 of 139	Official leaderboard
DeepSWE	53.3%	#20 of 28	Official leaderboard
Tencent WorkBuddy Bench	74.9%	#10 of 14	Official leaderboard
SWE-Bench Pro	52.6%	#60 of 81	Lab model card
FeatureBench	41.9%	#12 of 16	Official leaderboard
FeatureBench Lite	41.9%	#12 of 16	Official leaderboard
Monthly-SWEBench	45%	#9 of 12	Official leaderboard
ReviewBench	64%	#3 of 4	Official leaderboard
EvoCode-Bench	9.40%	#10 of 13	Official leaderboard
SWE-bench Science	23.5%	#8 of 10	Official leaderboard
AndroidBench	54.7%	#26 of 32	Official leaderboard
DesignArena Data Visualization	1,125	#98 of 118	Official leaderboard
SWE Refactor Bench	7.00%	#7 of 8	Official leaderboard

Agents and tools

Benchmark	Score	Rank	Source
Toolathlon	48.2%	#10 of 50	Official leaderboard
PlanBench-XL	63.1%	#2 of 9	Official leaderboard
PinchBench 2.0	81.7%	#14 of 57	Official leaderboard
LisanBench	2789.0	#36 of 133	Official leaderboard
GAIA	69.1%	#19 of 64	Official leaderboard
Long-Horizon-Terminal-Bench (LHTB)	45.5%	#8 of 25	Official leaderboard
ClawProBench	61.5%	#17 of 49	Official leaderboard
CAR-bench	60%	#5 of 12	Official leaderboard
AutomationBench	18.1%	#18 of 42	Official leaderboard
Terminal-Bench 2.0	56.9%	#28 of 63	Lab model card
Long-Horizon-Terminal-Bench 2.0	52.4%	#7 of 15	Official leaderboard
RuneBench	4.76	#38 of 76	Official leaderboard
AdaPlanBench	35.5%	#5 of 10	Official leaderboard
BrowseComp	73.2%	#44 of 81	Lab model card
BabaIsBench	75%	#18 of 32	Official leaderboard
BrowseComp Extended	22%	#24 of 42	Official leaderboard
Terminal-Bench 2.1 Extended	30%	#26 of 44	Official leaderboard
Claw-Eval Pass^3	57.8%	#15 of 24	Official leaderboard
MCP Atlas	69%	#31 of 46	Lab model card
CivBench	1,477	#13 of 19	Official leaderboard
SkillsBench	44.7%	#16 of 22	Official leaderboard
Claw-Eval	74.2%	#18 of 24	Official leaderboard
Toolathlon-Verified	50.9%	#19 of 25	Official leaderboard
LiveClawBench	60.5%	#13 of 17	Official leaderboard
ClawBench V2	2.30%	#6 of 7	Official leaderboard
PACT	1,483	#23 of 26	Official leaderboard
CLASHarness	8.50%	#14 of 14	Official leaderboard

Professional work

Benchmark	Score	Rank	Source
LEXam-hard	38.2%	#3 of 15	Official leaderboard
FORTE	70.7%	#9 of 21	Official leaderboard
APEX-Agents 1.1	67.5%	#31 of 65	Official leaderboard
APEX-1	63.8%	#37 of 71	Official leaderboard
CommerceAgentBench	46.7%	#8 of 13	Official leaderboard
FoodTruck Bench	$5,504	#41 of 59	Official leaderboard
APEX-Accounting	53.3%	#40 of 49	Official leaderboard
DAYJOB: Finance	0.30%	#34 of 39	Official leaderboard
MerchantBench	24.7	#7 of 8	Official leaderboard
Clipping Bench	$320	#17 of 19	Official leaderboard
GDPval-AA	1,395	#15 of 16	Lab model card
DAYJOB: Healthcare	0%	#35 of 37	Official leaderboard

Vision, audio and robotics

Benchmark	Score		Rank	Source
ScienceVision	0.27		#10 of 12	Official leaderboard

Safety and security

Benchmark	Score	Rank	Source
CyberGym	97.1%	#3 of 34	Official leaderboard
nocot-bench (NCRI)	78.1	#63 of 185	Official leaderboard
AILuminate Safety 1.2	43.0	#2 of 5	Official leaderboard
MLCommons AILuminate	43%	#2 of 5	Official leaderboard
DeepsecBench	16.4	#17 of 29	Official leaderboard
CyberBench (Vals)	66.7%	#16 of 26	Official leaderboard
SecCodeBench	54.7%	#5 of 6	Official leaderboard
dfbench	31.8%	#17 of 18	Official leaderboard

Other

Benchmark	Score	Rank	Source
Conceptual Reasoning Index	48.5	#75 of 151	Official leaderboard
LMCA	35.9%	#76 of 147	Official leaderboard
Political Science LLM Benchmark	66.5%	#22 of 42	Official leaderboard
ACCoRD	57.3	#97 of 152	Official leaderboard
CompassRank	73.7	#12 of 14	Official leaderboard

Its lab's numbers put it at 150.0, 3.1 points above its ECI

The Self-Reported Index fits one score to the 17 capability results DeepSeek reported for this model, on the scale of the Epoch Capabilities Index. Labs choose which benchmarks to show and how to run them, so read it as the lab's claim, not an independent measurement. The 90% interval runs from 148.2 to 151.8. How we compute it.

Its lab's numbers next to 4 leaderboards

Where the benchmark's own leaderboard, or Epoch AI's run, also has this model. Gaps mostly come from different harnesses, effort levels and task subsets. Every pair.

Benchmark	Lab says	Board says	Gap
MathArena (Apex, IMO, USAMO)	33% Apex (Pass@1), V4-Flash Max	27.1% Leaderboard, max	+5.9 pts
MathArena Apex Shortlist	85.7% Apex Shortlist (Pass@1), V4-Flash Max	89.4% Leaderboard, max	−3.7 pts
HMMT 2026	94.8% HMMT 2026 Feb (Pass@1), V4-Flash Max	93.9% Leaderboard, max	+0.9 pts
Toolathlon	47.8% Toolathlon (Pass@1), V4-Flash Max	48.2% Leaderboard, max	−0.4 pts

DeepSeek's models over time

DeepSeek's models in color, against every other model on the Epoch Capabilities Index, by release date.

Source: Epoch Capabilities Index, Epoch AI (CC BY 4.0)

Its launch post

The benchmarks the lab chose to headline, with the scores it gave.

DeepSeek-V4-Flash model card (Hugging Face) Apr 24, 2026

Its system card

Every result in the longer documents the lab published with it. 1 of them isn't in the launch post.

DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence (arXiv 2606.19348) Apr 26, 2026

Capabilities

Hallucination

SimpleQA Verified 34.1
Chinese SimpleQA 78.9