DeepSeek v3 — Benchmarks

Benchmark scores for DeepSeek v3 aggregated from public leaderboards (18 boards), with how it ranks among open models and how it compares with its nearest neighbours. See hardware requirements for what you need to run it.

llmrun Composite Score →

89.0

90% range 85.4–92.5 · from 8 boards

Coding96.990% range 91–103
AgentsNot enough data yet
MathNot enough data yet
Science89.590% range 82–97
Reasoning86.990% range 72–102

DeepSeek v3 reasoning benchmarks

Multi-step logic and problem-solving — can the model think through a hard problem rather than recall an answer.

GPQA Diamond
GPQA Diamond is a set of extremely hard, graduate-level science questions (physics, chemistry, biology) written by domain experts and filtered so that skilled non-experts with web access still fail. It measures genuine reasoning rather than memorization.
56.5%
56.5%
median 55.7%open leader 93.2%
Overall best: Gemini 4 Argon (high) 96.4%
54th percentile · #47 of 100 open · #219 of 309 all
#47 of 100
#219 of 309
Independent
Epoch AI · Dec 2024
SimpleBench
SimpleBench is a set of everyday, common-sense and trick questions that humans answer easily but language models often get wrong. It probes basic reasoning and robustness rather than specialist knowledge.
18.9%
18.9%
median 40.8%open leader 61.1%
Overall best: Claude Fable 5 Max 81.9%
4th percentile · #24 of 25 open · #96 of 101 all
#24 of 25
#96 of 101
Independent
Epoch AI · Dec 2024
CritPt
CritPt (Critical Points) tests graduate-level physics problem-solving on research-style questions spanning mechanics, electromagnetism, quantum mechanics, and more. It was created by physicists to probe whether a model can reason through multi-step physics rather than pattern-match textbook answers.
0.0%
0.0%
median 0.9%open leader 26.6%
Overall best: Gemini 4 Carbon (high) 36.0%
19th percentile · #54 of 84 open · #197 of 244 all
#54 of 84
#197 of 244
Lab-reported
Epoch AI · Dec 2024
ForecastBench
ForecastBench, from the Forecasting Research Institute, asks models to predict real-world future events, with questions that only resolve after the model's training cutoff so the answers can't have been memorised. It is scored as a difficulty-adjusted Brier index, where higher is better.
62.6
62.6
median 61.5open leader 64.7
Overall best: GPT 5.5 (high) 67.0
76th percentile · #6 of 24 open · #58 of 92 all
#6 of 24
#58 of 92
Lab-reported
Epoch AI · Dec 2024

Other results (2)leaderboard-only, not part of the llmrun Score

Leaderboard-only, not part of the llmrun Score

BIG-Bench Hard
BIG-Bench Hard (BBH) is the 23-task subset of BIG-Bench that earlier models found hardest — multi-step reasoning, logic, and word problems. It's a long-standing standard for tracking reasoning ability across model generations.
87.5%
87.5%
median 50.8%open leader
Overall best: Gemini 1.5 Pro 001 89.2%
top · #1 of 38 open · #2 of 50 all
#1 of 38
#2 of 50
Lab-reported
Epoch AI · Dec 2024
WinoGrande
WinoGrande is a large-scale commonsense-reasoning test built around pronoun resolution: the model must pick which of two nouns a pronoun refers to in a sentence, using real-world knowledge to disambiguate. It was designed by AI2 to resist the shortcuts that let earlier models game its predecessor, Winograd.
85.2%
85.2%
median 70.1%open leader 89.2%
94th percentile · #4 of 51 open · #7 of 79 all
#4 of 51
#7 of 79
Lab-reported
Epoch AI · Dec 2024

DeepSeek v3 coding benchmarks

Writing, editing, and fixing real code — measured by whether the resulting program actually runs and passes tests.

Aider Polyglot
The Aider Polyglot benchmark measures real-world coding across several programming languages: the model edits code to solve Exercism exercises, and is scored on whether the final solution actually runs and passes the tests.
48.4%
48.4%
median 45.1%open leader 74.2%
Overall best: GPT 5 (Aug 07, 2025, high) 88.0%
53rd percentile · #9 of 18 open · #40 of 69 all
#9 of 18
#40 of 69
Independent
Epoch AI · Dec 2024
SciCode
SciCode asks a model to write real scientific-computing code — implementing multi-step numerical methods drawn from physics, chemistry, biology, and math research — and checks the output against reference solutions. It was built by a team of scientists and ML researchers to test coding ability on genuine research problems, not typical software tasks.
35.8%
35.8%
median 40.1%open leader 60.9%
Overall best: Claude Opus 5.5 Max 66.9%
28th percentile · #52 of 72 open · #192 of 217 all
#52 of 72
#192 of 217
Lab-reported
Epoch AI · Dec 2024
SWE-bench Lite
SWE-bench Lite is a smaller, lower-cost subset of SWE-bench focused on self-contained bug fixes. It is the quickest of the SWE-bench boards to run and a common entry point for comparing coding agents.
36.7%
36.7%
median 36.7%open leader 49.7%
Overall best: ExpeRepair-v1.0 + Claude 4 Sonnet 60.3%
50th percentile · #2 of 3 open · #34 of 80 all
#2 of 3
#34 of 80
Independent
SWE-bench · Jun 2025

Other results (1)leaderboard-only, not part of the llmrun Score

Leaderboard-only, not part of the llmrun Score

METR Time Horizons
METR Time Horizons, from the AI-evaluation nonprofit METR, runs models on a suite of software-engineering and research tasks whose length is measured by how long they take skilled humans, from seconds to many hours. METR summarises each model as a time horizon: the human task length at which the model succeeds half the time. llmrun shows METR's average task-success score (0–1) as published by Epoch AI; higher means the model completes more of the suite, including longer tasks.
47.4%
47.4%
median 49.6%open leader 59.2%
Overall best: Claude Mythos Preview Early 85.2%
38th percentile · #6 of 9 open · #32 of 46 all
#6 of 9
#32 of 46
Lab-reported
Epoch AI · Dec 2024

DeepSeek v3 math benchmarks

Symbolic and quantitative problem-solving, from competition math to multi-step calculation.

AIME 2024/2025
AIME (American Invitational Mathematics Examination) is a prestigious high-school competition of hard, integer-answer problems. It's a widely-cited yardstick for multi-step mathematical reasoning, here on the 2024–2025 papers.
15.8%
15.8%
median 54.4%open leader 99.4%
Overall best: Claude Fable 5 (high) 100.0%
34th percentile · #58 of 88 open · #239 of 292 all
#58 of 88
#239 of 292
Independent
Epoch AI · Dec 2024
MATH Level 5
MATH Level 5 covers the hardest tier of competition-mathematics problems, testing multi-step symbolic and quantitative reasoning.
64.8%
64.8%
median 41.6%open leader 96.6%
Overall best: GPT 5 (Aug 07, 2025, high) 98.1%
76th percentile · #10 of 39 open · #50 of 107 all
#10 of 39
#50 of 107
Independent
Epoch AI · Dec 2024

DeepSeek v3 knowledge benchmarks

Breadth and depth of factual knowledge across many subjects, paired with the reasoning to apply it.

MMLU-Pro
MMLU-Pro is a harder, cleaned-up successor to MMLU with ten answer choices and more reasoning-heavy questions across 14 subjects, measuring broad knowledge and reasoning together.
75.9%
75.9%
median 52.0%open leader 88.0%
Overall best: Gemini-3.1-Pro 91.2%
81st percentile · #29 of 146 open · #82 of 257 all
#29 of 146
#82 of 257
Independent
TIGER-Lab

Other results (4)leaderboard-only, not part of the llmrun Score

Leaderboard-only, not part of the llmrun Score

MMLU
MMLU (Massive Multitask Language Understanding) spans 57 subjects from history to law to medicine as multiple-choice questions. It is the long-standing default for broad knowledge and remains the most widely-reported general benchmark.
87.2%
87.2%
median 62.5%open leader
Overall best: GPT 4o (Nov 20, 2024) 88.1%
top · #1 of 85 open · #3 of 136 all
#1 of 85
#3 of 136
Lab-reported
Epoch AI · Dec 2024
ARC Challenge
ARC Challenge is a set of grade-school science questions, curated by the Allen Institute for AI (AI2) to exclude ones that simple word-matching or retrieval can answer. It measures science knowledge paired with the reasoning needed to apply it.
95.3%
95.3%
median 53.9%open leader
top · #1 of 56 open · #1 of 76 all
#1 of 56
#1 of 76
Lab-reported
Epoch AI · Dec 2024
HellaSwag
HellaSwag is a commonsense benchmark that asks a model to pick the most plausible continuation of an everyday situation. The wrong options are adversarially chosen to fool models, testing grounded commonsense understanding.
88.9%
88.9%
median 76.2%open leader 89.2%
Overall best: GPT 4 (Mar 14) 95.3%
96th percentile · #3 of 47 open · #5 of 75 all
#3 of 47
#5 of 75
Lab-reported
Epoch AI · Dec 2024
TriviaQA
TriviaQA is a large set of trivia questions paired with evidence documents, testing a model's factual recall and reading comprehension across a wide range of everyday topics.
82.9%
82.9%
median 73.7%open leader 87.6%
94th percentile · #2 of 19 open · #10 of 39 all
#2 of 19
#10 of 39
Lab-reported
Epoch AI · Dec 2024

DeepSeek v3 instruction following benchmarks

How precisely a model obeys explicit, verifiable instructions — format, length, and content constraints — and, more broadly, how much people prefer its responses head-to-head in blind human voting.

LMArena Text
Arena (arena.ai, formerly LMArena / Chatbot Arena) ranks models by millions of blind, head-to-head human votes on real prompts. The Arena Score shown here is a Bradley-Terry rating computed with Arena's default style control, so higher means people preferred that model's answers more often — it's a relative rating, not a percentage, so compare it only to other models on this board.
1358.4
1358.4
median 1321.4open leader 1488.2
Overall best: Gemini 4 Carbon (High) 1527.5
63rd percentile · #71 of 188 open · #202 of 408 all
#71 of 188
#202 of 408
Arena
Arena · Oct 2026

Score vs VRAM in the ≤512 GB tier

DeepSeek v3
Other models
Best Score for its VRAM
x: VRAM (GB, log scale) · y: llmrun Score

#72 of 102 that fit ≤512 GB

Each dot is a model at the best quant that fits ≤512 GB. The line steps up to the best Score available at each amount of VRAM; a dot below it is beaten by a model that needs no more memory.

Full ≤512 GB ranking →

DeepSeek v3 vs its neighbours

Same boards for the nearest open models per llmrun Score. The highest score in each row is marked.

Board	DeepSeek v3this model	Qwen3.8 27Bbest in classQ4_K_M · 18 GB	DeepSeek v3 0324nearest by ScoreQ4_K_M · 412 GB	Mistral Large 3 675B Instruct 2512nearest by ScoreQ4_K_M · 454 GB	Llama 4 Maverick 17B 128E Instructnearest by ScoreQ4_K_M · 273 GB
llmrun Score	89.0	122.0	93.9	89.6	87.1
GPQA Diamond	56.5%	—	67.6%	—	67.0%
SimpleBench	18.9%	—	27.2%	—	27.7%
CritPt	0.0%	5.4%	0.0%	0.0%	0.0%
Aider Polyglot	48.4%	—	55.1%	—	15.6%
SciCode	35.8%	46.6%	39.0%	36.6%	31.7%
AIME 2024/2025	15.8%	—	37.8%	—	20.6%
MATH Level 5	64.8%	—	75.5%	—	73.0%
MMLU-Pro	75.9%	—	81.3%	—	80.5%
LMArena Text	1358.4	1435.6	1395.2	1414.3	1327.0
Compare head to head		Compare	Compare	Compare	Compare

Scores aggregated from public benchmark sources (Epoch AI: 15 boards, Arena: 1 board, SWE-bench: 1 board, TIGER-Lab: 1 board) — each linked from the benchmark pages. llmrun does not run these benchmarks. Percentiles and ranks are among open models in the llmrun catalogue; the median and leader marks refer to the same field.