DE
DeepSeek
Language model · China ·Dec 2024

DeepSeek-V3

Language Open (Restricted)
671B
Parameters
45
Benchmarks

About DeepSeek-V3

DeepSeek-V3 is an AI model developed by DeepSeek, in the language category, released in 2024, made available as an open-weights model with 671B params.

On this page you'll find DeepSeek-V3's full specifications, including 45 benchmark results. Review provider pricing and benchmark scores below, or compare DeepSeek-V3 head-to-head with other language models.

Links

Run DeepSeek-V3 locally — what it takes

DeepSeek-V3 (671B) is too large to run on a single consumer or Apple machine. At Q4 its weights alone are about 376 GB — it needs a multi-GPU / datacenter setup:

  • ≈ 2× AMD Instinct MI325X
  • ≈ 3× NVIDIA B200

Lower quantization (Q3/Q2) or CPU + system-RAM offload (e.g. ktransformers) reduces the requirement at the cost of speed.

Software support

✓ measured · · compatible · — not supported. Informational only — speed is hardware-based.

Community

Would you run DeepSeek-V3 again?
No votes yet — be the first.
Did it run? — community reports by hardware

No reports yet. Own this model on some hardware? Be the first to confirm it runs.

Discussion

No comments yet. Start the discussion.

DeepSeek-V3 benchmark scores
Benchmark Category Score Variant
Aider Polyglot code 49.6% cited-magistral
HumanEval code 94.6% pass@10-cited-tulu3
HumanEval code 92.1% cited-minimax-text
HumanEval code 91.5% pass@1
HumanEval+ code 91.6% pass@10-cited-tulu3
LiveCodeBench code 42.7% pass@1
LiveCodeBench v5 code 36.2% cited-magistral
MBPP+ code 78.8% plus-cited-minimax-text
MBPP+ code 75.7% pass@1
AlpacaEval 2 general 66.5%
AlpacaEval 2 general 53.5% lc-win-cited-tulu3
ArenaHard general 92.1%
ArenaHard general 91.4% cited-minimax-text
C-Eval general 86.3%
CLUEWSC general 93.5%
IFEval general 88% cited-tulu3
IFEval general 87.3% avg-cited-minimax-text
IFEval general 86.1%
MMLU general 88.5% cited-minimax-text
MMLU general 87.9% EM
MMLU general 82.1% cited-tulu3
MMLU-Pro general 76% EM
MMLU-Pro general 75.9% cited-minimax-text
MMLU-Redux general 88.8% EM
AIME 2024 math 39.2% cited-magistral
AIME 2024 math 34% pass@1
AIME 2025 math 28.8% cited-magistral
CNMO 2024 math 33.9% pass@1
GSM8K math 96.7% cited-minimax-text
GSM8K math 94.1% cited-tulu3
MATH math 89.7% EM
MATH math 84.6% cited-minimax-text
MATH math 72.5% cited-tulu3
MATH-500 math 90.2% cited-magistral
MATH-500 math 88.9% pass@1
BIG-Bench Hard reasoning 89.5% cited-tulu3
DROP reasoning 91.8% F1
DROP reasoning 91% f1-cited-minimax-text
GPQA Diamond reasoning 59.1% cited-magistral
GPQA Diamond reasoning 59.1% diamond-cited-minimax-text
GPQA Diamond reasoning 56.1% pass@1
C-SimpleQA safety 68.9%
C-SimpleQA safety 64.8% cited-minimax-text
SimpleQA safety 24.9% cited-minimax-text
SimpleQA safety 24.6%

More models like DeepSeek-V3

Frequently asked questions

What is DeepSeek-V3?

DeepSeek-V3 is an AI model developed by DeepSeek, in the language category, released in 2024. It is tracked on GenAIList with its specifications, benchmark scores and provider pricing.

Who created DeepSeek-V3?

DeepSeek-V3 was developed by DeepSeek and released in 2024.

Is DeepSeek-V3 open source or proprietary?

DeepSeek-V3 ships as an open-weights model: you can download and self-host the weights, though the licence may place some restrictions on use.

How much does DeepSeek-V3 cost?

Pricing for DeepSeek-V3 depends on the provider. See the providers table on this page for the latest API rates.

How does DeepSeek-V3 perform on benchmarks?

DeepSeek-V3 is benchmarked across 45 evaluations on GenAIList, including Aider Polyglot (49.6%). See the full benchmark table below and compare it with other models.

No videos yet. Know a good one? Add it below.

Know a great video about DeepSeek-V3?