DeepSeek V3 0324, DeepSeek's 03 March 2025 V3 model, optimized for general-purpose tasks. Quantized at FP8.
Context Window
128.0K
Max Output
8.2K
Avg output tokens (7d)
420 tokens
Input Price (Auto)
$0.20/1M
Output Price (Auto)
$0.77/1M
Cache Read (Auto)
$0.14/1M
Capabilities
Benchmarks
Benchmarks
Performance metrics and benchmarks
Sourced from Artificial Analysis.
Intelligence Index
9.7
Coding Index
21.2
Agentic Index
0.8
Agentic work
AutomationBench-AA
Workflow automation with guardrail penalties
0.3%
Better than 6% of models compared
AA-Briefcase
Agentic knowledge work (Elo)
92 Elo
Better than 7% of models compared
GDPval-AA v2
Economically valuable tasks (Elo)
275 Elo
Better than 12% of models compared
Document reasoning
GDP.pdf
Professional PDF reasoning: all-pass rate
1.6%
Better than 17% of models compared
AA-LCR v1.1
Long context reasoning with updated grading
40.7%
Better than 39% of models compared
Reasoning
HLE
Humanity's Last Exam
4.7%
Better than 26% of models compared
IFBench
Instruction-following benchmark
41.0%
Better than 40% of models compared
CritPt
Research-level physics reasoning
0.0%
Coding
Terminal-Bench v4.0
Practical coding and terminal tasks
0.0%
Better than 17% of models compared
SciCode
Python programming for scientific computing
39.0%
Better than 25% of models compared
LiveCodeBench
Contamination-free coding benchmark
40.5%
Better than 48% of models compared
Math
AIME 2025
American Invitational Mathematics Examination 2025
41.0%
Better than 42% of models compared
AIME
American Invitational Mathematics Examination
52.0%
Better than 72% of models compared
Math-500
Diverse mathematical problem solving benchmark
94.2%
Better than 76% of models compared
Knowledge
MMLU-Pro
Professional and academic subject knowledge
81.9%
Better than 77% of models compared
AA-Omniscience Accuracy
Proportion of correctly answered questions
24.3%
AA-Omniscience Hallucination Rate
Rate of incorrect answers among non-correct responses
85.9%
Legacy benchmarks
GPQA Diamond (legacy)
Graduate-level scientific reasoning
65.5%
Better than 41% of models compared
Terminal-Bench Hard (legacy)
Agentic coding and terminal use
15.2%
Better than 52% of models compared
T²-Bench Telecom (legacy)
Conversational AI agents in dual-control scenarios
47.1%
Better than 51% of models compared
AA-LCR (unversioned / legacy)
Long context reasoning evaluation
40.7%
Better than 39% of models compared
GDPval-AA (unversioned / legacy)
Economically valuable tasks
0.0%
Last updated Sep 13, 2026
Artificial AnalysisProviders
Choose explicit providers for this model. Auto routing remains available as the default option.
Loading provider options…
Related text models
Compare DeepSeek Chat 0324 with similar models from the same provider or model family.
DeepSeek V3/Chat Cheaper
deepseek-chat-cheaperCheaper version of Deepseek V3/Chat. Note: may be routed through Deepseek itself. Quantized at FP8.
DeepSeek V3/Deepseek Chat
deepseek-chatDeepSeek original V3 model, trained on nearly 15 trillion tokens, matches leading closed-source models at a far lower price. Quantized at FP8.
DeepSeek V4.1 Flash TEE
TEE/deepseek-v4.1-flashDeepSeek V4.1 Flash supports text and image input, reasoning, tool calling, and structured output with a 1M-token context window. This route runs through Tinfoil attested inference inside a Trusted Execution Environment.
DeepSeek V4.1 Flash
deepseek/deepseek-v4.1-flashDeepSeek V4.1 Flash supports text and image input, reasoning, tool calling, and structured output with a 1M-token context window. This is a rate-limited beta with limited capacity, intended for testing rather than production use.
DeepSeek V4.1 Flash Thinking
deepseek/deepseek-v4.1-flash:thinkingDeepSeek V4.1 Flash supports text and image input, reasoning, tool calling, and structured output with a 1M-token context window. This is a rate-limited beta with limited capacity, intended for testing rather than production use.
DeepSeek V4 Flash Vision Exp
deepseek/deepseek-v4-flash-vision-expAn experimental vision-enabled DeepSeek V4 Flash model that adds image understanding while retaining the text, reasoning, coding, tool-calling, and agent capabilities of the base model.