DeepSeek-V3.1

DeepSeek-V3.1 places in the top three on 4 benchmarks.

DeepSeek Released August 21st, 2025 Family DeepSeek-V3 The lab's model card Compare with others

Benchmarks with a score

Ranked first, out of five or more models

In the top three

DeepSeek Aug 2025

DeepSeek-V3.1

Open weights Flagship

140.8 #120 of 256

Epoch Capabilities Index

Self-Reported Index 143.0 #154 of 404

MultiChallenge 46.1% #31 /42
BrowseComp-ZH 49.2% #16 /19
MathArena (Apex, IMO, USAMO) 0.52% #43 /47

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
SimpleQA	93.4%	#4 of 60	Lab model card
Vectara Hallucination Leaderboard (2025 dataset)	5.50%	#19 of 104	Official leaderboard
MMLU-Redux	93.7%	#9 of 45	Lab model card
SuperGPQA	59.3%	#12 of 59	Official leaderboard
MMLU-Pro	84.8%	#34 of 167	Lab model card
LMArena (Elo)	1,417	#108 of 335	Official leaderboard
ECLeKTic	19%	#16 of 35	Official leaderboard
Creative Writing v3	69.9	#56 of 120	Official leaderboard
HELM Arabic Enterprise	65.9	#11 of 23	Official leaderboard
Longform Creative Writing (EQ-Bench)	62.3	#59 of 119	Official leaderboard
Humanity's Last Exam	29.8%	#82 of 154	Lab model card
Phare	64.8%	#34 of 62	Official leaderboard
HELM Arabic	77.9	#16 of 28	Official leaderboard
Meeseeks Utility Rate	79.1%	#6 of 10	Official leaderboard
Wer wird Millionär? LLM Benchmark	1111.0	#48 of 77	Official leaderboard
SimpleQA Verified	24.3%	#36 of 56	Official leaderboard
MultiNRC	23.6%	#26 of 38	Official leaderboard
Vectara Hallucination Leaderboard (HHEM)	5.50%	#102 of 148	Official leaderboard
Fiction.LiveBench	52.8%	#40 of 57	Fiction.live leaderboard, via Epoch AI
MultiChallenge	46.1%	#31 of 42	Official leaderboard
Global MMLU Lite	80.4%	#32 of 43	Official leaderboard
Role-play Benchmark (MiniMax)	64.2	#6 of 8	Official leaderboard

Math and science

Benchmark	Score	Rank	Source
AIME 2024	93.1%	#7 of 88	Lab model card
AMO-Bench	47.6%	#3 of 21	Official leaderboard
IneqMath	15.5%	#8 of 34	Official leaderboard
GPQA Diamond	80.1%	#118 of 302	Lab model card
AIME 2025	88.4%	#58 of 143	Lab model card
DTBench	82.7%	#78 of 180	Official leaderboard
HMMT 2025	84.2%	#32 of 54	Lab model card
BRUMO 2025	90%	#25 of 42	Official leaderboard
SMT 2025	84%	#24 of 40	Official leaderboard
CMIMC 2025	82.5%	#20 of 33	Official leaderboard
CORE-Bench	20%	#12 of 17	Official leaderboard
SimpleBench	40%	#75 of 106	Official leaderboard
PHYBench	16.6%	#21 of 28	Official leaderboard
General365	48.1%	#19 of 23	Official leaderboard
ForecastBench	0.16	#69 of 82	Official leaderboard
MathArena (Apex, IMO, USAMO)	0.52%	#43 of 47	Official leaderboard

Coding

Benchmark	Score	Rank	Source
Aider Polyglot	76.3%	#5 of 30	Lab model card
CATArena Gomoku	1,456	#2 of 5	Official leaderboard
Codeforces rating	2,091	#20 of 47	Lab model card
LiveCodeBench v6	74.8%	#44 of 90	Lab model card
DesignArena UI Component	1,173	#79 of 139	Official leaderboard
DesignArena Website	1,180	#86 of 151	Official leaderboard
DesignArena 3D Design	1,123	#81 of 136	Official leaderboard
CATArena Texas Holdem	981	#3 of 5	Official leaderboard
DesignArena Code	1,166	#86 of 143	Official leaderboard
SWE-bench Verified	66%	#97 of 155	Lab model card
DesignArena Data Visualization	1,166	#80 of 118	Official leaderboard
DesignArena Game Dev	1,121	#99 of 143	Official leaderboard
SWE-fficiency	4.32%	#12 of 17	Official leaderboard
WeirdML	38.4%	#85 of 112	Official leaderboard
DesignArena SVG	1,055	#82 of 103	Official leaderboard
CATArena SWE-Perf-10	0.19×	#4 of 5	Official leaderboard
SWE-bench Multilingual	54.5%	#67 of 73	Lab model card
CATArena OIBench-5	0.04	#5 of 5	Official leaderboard

Agents and tools

Benchmark	Score	Rank	Source
Step Race	4	#12 of 59	Official leaderboard
Meta Agents Research Environments (ARE) / Gaia2	21.9%	#5 of 13	Official leaderboard
C3-Bench	8.98%	#12 of 30	Official leaderboard
Agent Poker Bench	1,058	#7 of 17	Official leaderboard
CompileBench	80%	#9 of 20	Official leaderboard
SnakeBench	23	#100 of 188	Official leaderboard
MCP-Universe	22.1%	#21 of 30	Official leaderboard
MCPMark	16.7%	#20 of 28	Official leaderboard
GAIA	27.9%	#46 of 64	Official leaderboard
BrowseComp-ZH	49.2%	#16 of 19	Lab model card
Terminal-Bench	31.3%	#19 of 22	Lab model card
BrowseComp	30%	#77 of 81	Lab model card

Professional work

Benchmark	Score	Rank	Source
ProfBench	46.6%	#38 of 72	Official leaderboard
StockBench	8.00	#8 of 13	Official leaderboard
CaseLaw v2	53.9%	#39 of 53	Official leaderboard
CorpFin v2	51.5%	#92 of 119	Official leaderboard
PRBench	37.6%	#42 of 48	Official leaderboard
FailureSensorIQ	36.6%	#23 of 24	Official leaderboard

Vision, audio and robotics

Benchmark	Score		Rank	Source
OmniVideoBench	27.6%		#8 of 8	Official leaderboard

Safety and security

Benchmark	Score	Rank	Source
SUPERChem	37.3%	#2 of 6	Official leaderboard
BrowseSafe++	75.2%	#10 of 15	Official leaderboard
Lakera B3 (Backbone Breaker Benchmark)	50.3%	#25 of 35	Official leaderboard
Human Pathogen Capabilities Test	41.6%	#65 of 86	Official leaderboard
Human Pathogen Capabilities Test v2 (HPCT v2)	39.8%	#56 of 71	Official leaderboard
World-Class Bio	28.6%	#64 of 74	Official leaderboard
Molecular Biology Capabilities Test	39.1%	#78 of 90	Official leaderboard
FORTRESS	60.5%	#69 of 73	Official leaderboard
BrowseSafe-Adapt	95%	#6 of 6	Official leaderboard

Other

Benchmark	Score	Rank	Source
Conceptual Reasoning Index	39.9	#100 of 151	Official leaderboard
ACCoRD	55.5	#105 of 152	Official leaderboard
LMCA	24.3%	#104 of 147	Official leaderboard

Its lab's numbers put it at 143.0, 2.1 points above its ECI

The Self-Reported Index fits one score to the 15 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 139.3 to 146.7. How we compute it.

Its lab's numbers next to 3 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
AIME 2025	88.4% AIME 2025 (Pass@1)	90.8% Leaderboard, thinking	−2.4 pts
HMMT 2025	84.2% HMMT 2025 (Pass@1)	85.8% Leaderboard, thinking	−1.6 pts
MMLU-Pro	84.8% MMLU-Pro (EM)	84.8% Leaderboard	Same

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-V3.1 model card Aug 21, 2025

MMLU-Redux (EM) 93.7
MMLU-Pro (EM) 84.8
GPQA-Diamond (Pass@1) 80.1
Best of Humanity's Last Exam (Pass@1) / Humanity's Last Exam (Python + Search) 29.8
BrowseComp 30
BrowseComp_zh 49.2
SimpleQA 93.4
LiveCodeBench (2408-2505) (Pass@1) 74.8
Codeforces-Div1 (Rating) 2091
Aider-Polyglot (Acc.) 76.3
SWE Verified (Agent mode) 66
SWE-bench Multilingual (Agent mode) 54.5
Terminal-bench (Terminus 1 framework) 31.3
AIME 2024 (Pass@1) 93.1
AIME 2025 (Pass@1) 88.4
HMMT 2025 (Pass@1) 84.2