DeepSeek V3 0324

  • Open weights
  • Standard
  • Input: text
  • Released 2025-03
RankedEnough results for an index.
IndexAll categories, out of 100.37.1±5.0Median 50.7
CoverageIndex weight that has results.90%
SpeedOutput tokens per second.—Median 76/s
First tokenSeconds to the first output token.—Median 1.1 s
Input / 1MUS dollars per 1M input tokens.—Median $0.60
Output / 1MUS dollars per 1M output tokens.—Median $2.50
ContextMaximum tokens in one request.128KMedian 262K
EloLMArena rating and rank.1375±4 #189Median 1425

Run locally

684.53B parameters
16-bit
1438.5 GB
8-bit
764.7 GB
4-bit
436.7 GB
Weights
Hugging Face (opens in a new tab) ↗

Memory for a short context. 4-bit needs more than 128 GB. Compare local models

Capabilities

Index per category, out of 100.
Out of 100
AgenticMulti-step tasks with tools.2 counted34.4193 / 221
CodingCode writing and repair.4 counted40.6184 / 214
ReasoningLogic problems and puzzles.2 counted35.9182 / 225
MultimodalTasks with images and text.None counted——
KnowledgeFacts and expert knowledge.4 counted37.8209 / 266
MultilingualTasks in many languages.1 counted40.1—
InstructionTasks with strict prompt rules.1 counted27.8—
MathMath problems.2 counted37.7133 / 171

Results

16 counted
BenchmarkResultThe score from the source.PlacePlace among results on this benchmark.
MATH 500Math88.6% ±1.430 / 59
AIMEMath52.2% ±1.760 / 95
Artificial Analysis IFBenchInstruction41.0%96 / 140
MMLU ProKnowledge79.5% ±0.494 / 135
Critical Physics TasksReasoning0.0%145 / 207
LiveCodeBenchCoding65.5% ±1.2100 / 142
Artificial Analysis GPQA DiamondKnowledge65.5%138 / 193
GDPval-AA normalizedAgentic0.0%100 / 132
Artificial Analysis SciCodeCoding39.0%90 / 113
GPQA Diamond · Vals AIKnowledge61.6% ±2.4109 / 135
MMLU-ProXMultilingual70.5%14 / 16
IOI v1Coding1.7% ±0.957 / 61
MGSMMultilingual91.7% ±1.726 / 74
TaxEval v2Finance71.1% ±0.983 / 143
CorpFin v2Finance54.7% ±1.092 / 131
MedQAHealthcare82.0% ±0.468 / 95
LegalBenchLegal77.7% ±0.4107 / 149

Sources

BenchLMCC BY-NC 4.0Vals AINo licence statedSWE-bench teamNo licence stated

History

13 releases
Index37.9 → 37.1
23 Oct 202621 Nov 2026

Questions

How good is DeepSeek V3 0324?
Its index is 37.1 out of 100. That is place 225 of 282 ranked models.
What is DeepSeek V3 0324 best at?
Math: 37.7 out of 100, place 133 of 171.
What is the context window of DeepSeek V3 0324?
128K tokens.
Is DeepSeek V3 0324 open source?
Yes. Its weights are published.
When was DeepSeek V3 0324 released?
March 2025.

Badge

DeepSeek V3 0324 on Inference360: place 225 of 282, index 37.1

Compare with

Ling 2.6 Flash37.1MiniCPM5-2B37.0Nemotron 3.5 Lightning 30B A3B NVFP436.8Granite 4.2 8B37.5DeepSeek-R138.8

More from DeepSeek

All models
DeepSeek-R138.8DeepSeek R1 Distill Llama 70BUnrankedDeepSeek R1 Distill Qwen 14BUnrankedDeepSeek R1 Distill Qwen 32BUnrankedDeepSeek-R1-052844.5