Inference
EN
DeepSeek

DeepSeek V3

DeepSeek V3 is a standard model from DeepSeek in the DeepSeek family. 22 benchmarks count toward its index, in 6 categories.

  • Open weights
  • Standard
  • Input: text
  • Released 2024-12
RankedEnough results for an index.
IndexAll categories, out of 100.29.9±4.9Median 51.0
CoverageIndex weight that has results.85%
SpeedOutput tokens per second.37/sMedian 67/s
First tokenSeconds to the first output token.1.3 sMedian 1.3 s
Input / 1MUS dollars per 1M input tokens.$0.27Median $0.60
Output / 1MUS dollars per 1M output tokens.$1.10Median $2.55
ContextMaximum tokens in one request.128KMedian 262K
EloLMArena rating and rank.1333±5 #229Median 1422

Run locally

684.5B parameters
16-bit
1438.5 GB
8-bit
764.6 GB
4-bit
436.7 GB
Weights
Hugging Face (opens in a new tab) ↗

Memory need with a short context. The 4-bit copy needs more than 128 GB. Compare local models

Capabilities

Index per category, out of 100.
Out of 100
AgenticMulti-step tasks with tools.2 counted26.0211 / 218
CodingCode writing and repair.4 counted31.1197 / 209
ReasoningLogic problems and puzzles.3 counted34.1195 / 227
MultimodalTasks with images and text.None counted——
KnowledgeFacts and expert knowledge.5 counted34.3228 / 268
MultilingualTasks in many languages.None counted——
InstructionTasks with strict prompt rules.2 counted29.153 / 63
MathMath problems.6 counted27.2149 / 168

Results

22 counted
BenchmarkThe benchmark name.ResultThe score from the source.PlacePlace among results on this benchmark.
SWE-bench Lite · SWE-bench/experimentsCoding36.7%5 / 15
SWE-bench Lite · swebench.comCoding30.7%6 / 15
MATH level 5Math64.9% ±1.037 / 88
SWE-bench Lite · SWE-bench/experimentsCoding30.7%7 / 15
Instruction-Following EvalInstruction86.1%36 / 70
Measuring Short-Form Factuality in Large Language ModelsKnowledge24.9%8 / 14
GPQA diamond · Epoch AIKnowledge56.5% ±2.8131 / 217
Artificial Analysis Omniscience AccuracyKnowledge25.5%103 / 164
Massive Multitask Language Understanding ProfessionalKnowledge75.9%48 / 75
MATH 500Math80.4% ±1.838 / 59
Critical Physics TasksReasoning0.0%150 / 214
OTIS Mock AIME 2024-2025Math15.8% ±4.3150 / 200
GDPval-AA normalizedAgentic0.0%101 / 133
MATH-500 Problem SetMath90.2%21 / 27
AIMEMath27.5% ±0.774 / 95
Graduate-Level Google-Proof Q&AKnowledge59.1%112 / 141
FrontierMath-2025-02-28-PrivateMath1.7% ±0.855 / 69
Artificial Analysis IFBenchInstruction34.8%122 / 146
MMLU ProKnowledge73.8% ±0.4113 / 135
Artificial Analysis Long Context ReasoningReasoning29.3%183 / 217
Artificial Analysis GPQA DiamondKnowledge55.7%168 / 199
τ²-Bench Tool-Agent-User EvaluationAgentic22.8%126 / 149
GPQA Diamond · Vals AIKnowledge54.5% ±2.5116 / 135
Artificial Analysis SciCodeCoding35.8%100 / 114
LongBench v2Reasoning48.7%22 / 25
American Invitational Mathematics Examination 2024Math39.2%24 / 27
Software Engineering Benchmark VerifiedCoding42.0%107 / 114
Artificial Analysis Humanity's Last ExamKnowledge2.9%220 / 222
LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for CodeCoding37.6%12 / 12
MGSMMultilingual92.1% ±1.721 / 74
FrontierMath-2025-02-28-PublicMath0.0%33 / 64
LegalBenchLegal80.8% ±0.483 / 149
TaxEval v2Finance67.9% ±0.9104 / 143
MedQAHealthcare80.9% ±0.471 / 95
CorpFin v2Finance52.5% ±1.0101 / 131

Sources

BenchLMCC BY-NC 4.0Vals AINo licence statedEpoch AICC BYSWE-bench teamNo licence stated

History

5 releases
Index30.4 → 29.9
23 Oct 202612 Nov 2026
Output / 1M$1.10 → $1.10
23 Oct 202612 Nov 2026

Questions

How much does DeepSeek V3 cost?
$0.27 per 1M input tokens, $1.10 per 1M output tokens.
How good is DeepSeek V3?
Its index is 29.9 out of 100. That is place 250 of 283 ranked models.
What is DeepSeek V3 best at?
Instruction: 29.1 out of 100, place 53 of 63.
How fast is DeepSeek V3?
37 output tokens per second, 1.3 s to the first token.
What is the context window of DeepSeek V3?
128K tokens.
Is DeepSeek V3 open source?
Yes. Its weights are published.
When was DeepSeek V3 released?
December 2024.

Badge

DeepSeek V3 on Inference360: place 250 of 283, index 29.9
HTML
<a href="https://inference360.ai/models/deepseek-v3"><img src="https://inference360.ai/badges/deepseek-v3.svg" alt="DeepSeek V3 on Inference360: place 250 of 283, index 29.9" height="20"></a>
Markdown
[![DeepSeek V3 on Inference360: place 250 of 283, index 29.9](https://inference360.ai/badges/deepseek-v3.svg)](https://inference360.ai/models/deepseek-v3)

Compare with

Sarvam 30B30.3Gemini 2.0 Flash29.5LFM2.5-2.6B29.3LFM2.5-8B-A1B30.6DeepSeek-R138.9

Lower price

dots3-note Preview63.4 · FreeApodex 1.1 Mini55.7 · FreeNemotron 3 Ultra50.8 · Free

More from DeepSeek

All models
DeepSeek-R138.9DeepSeek R1 Distill Qwen 32BUnrankedR1 Distill Llama 70BUnrankedDeepSeek V3 032437.1DeepSeek-V2.5Unranked