DeepSeek·released 2024-12-25✓ 2 sources
DeepSeek-V3 benchmark scores: 19 benchmarks tracked, leading 1. 21% of its scores are independently reproduced — the rest are self-reported or unverified.
Leads on 1 of 18 benchmarks·best result 1134 on Codeforces rating (self-reported)·4 of 19 independently reproduced·$0.27/$1.1 per M tokens
Head-to-headDeepSeek-V3 vs Llama 3.1-405BConsensus: LiteLLM · See every model’s pricing →
Leads on: ARC AI2
| Benchmark | Score | Evidence status | Measured |
|---|---|---|---|
| Codeforces rating | 1134rating | self-reported· optimizedT1 | 2025-01-22DeepSeek AI |
| ARC AI2 | 93.73accuracy (%) | self-reported· optimizedT1 | 2024-12-26DeepSeek-V3 Technical Report |
| MATH-500 | 90.2pass@1 (%) | self-reported· optimizedT1 | 2025-01-22DeepSeek AI |
| HellaSwag | 85.2accuracy (%) | self-reported· optimizedT1 | 2024-12-26DeepSeek-V3 Technical Report |
| BBH | 83.33accuracy (%) | self-reported· optimizedT1 | 2024-12-26DeepSeek-V3 Technical Report |
| MMLU | 82.93accuracy (%) | unverified· optimizedT1 | 2024-12-26Stanford CRFM Leaderboard |
| TriviaQA | 82.9accuracy (%) | self-reported· optimizedT1 | 2024-12-26DeepSeek-V3 Technical Report |
| Winogrande | 70.4accuracy (%) | self-reported· optimizedT1 | 2024-12-26DeepSeek-V3 Technical Report |
| PIQA | 69.4accuracy (%) | self-reported· optimizedT1 | 2024-12-26DeepSeek-V3 Technical Report |
| MATH level 5 | 64.85accuracy (%) | reproduced· optimizedT1 | 2024-12-26Epoch AI |
| GPQA diamond | 59.1accuracy (%) | self-reported· optimizedT1 | 2025-01-22DeepSeek AI |
| Aider polyglot | 48.4accuracy (%) | unverified· optimizedT1 | 2024-12-26Aider LLM Leaderboards |
| METR Time Horizons | 47.36accuracy (%) | unverifiedT1 | 2024-12-26METR - Measuring AI Ability to Complete Long Tasks |
| GPQA diamond | 42.05accuracy (%) | reproduced· optimizedT1 | 2024-12-26Epoch AI |
| AIME | 39.2pass@1 (%) | self-reported· optimizedT1 | 2025-01-22DeepSeek AI |
| LiveCodeBench | 36.2pass@1 (%) | self-reported· optimizedT1 | 2025-01-22DeepSeek AI |
| OTIS Mock AIME 2024-2025 | 15.75accuracy (%) | reproduced· optimizedT1 | 2024-12-26Epoch AI |
| FrontierMath-2025-02-28-Private | 3.02accuracy (%) | reproduced· optimizedT1 | 2024-12-26Epoch AI |
| SimpleBench | 2.68accuracy (%) | unverifiedT1 | 2024-12-26SimpleBench Leaderboard |
Claims drawn from cited facts, not live model generation.
Developed by DeepSeek and released in 2024, this model is a large, frontier-scale system with 684.53 billion parameters. That scale buys substantial capacity and headroom but comes with higher compute and cost than compact or mid-size alternatives.
4 cited facts
This model is tracked on 18 benchmarks. Of those, it currently holds the top score on 1. Because at least one claimed result conflicts with an independent measurement, its self-reported numbers should be treated with real caution.
4 cited facts
On average it trails the leading score by 76.25 points, a wide gap that reads as well behind the leaders rather than at the frontier or merely competitive. It holds the top score on 1 benchmark, so that is isolated category leadership on one benchmark, not broad front-of-pack standing.
2 cited facts
Benchmarks with more than one recorded measurement for this model — every one shown, not just the headline number.
7 facts cross-checked across data sources: 1 corroborated, 6 single-source.
Source facts, citations, and refresh stamp for this record.
Sources: curated_benchmark_scores, curated_capability_claims, epoch_benchmarks, huggingface_models +1 more
DeepSeek-V3 is an AI model developed by DeepSeek, released 2024-12-25. Its benchmark record below tags every score as independently reproduced, self-reported, or unverified.
DeepSeek-V3 has recorded scores on 18 benchmarks, each shown with its evidence status.
DeepSeek-V3 has recorded scores on 19 benchmarks — PIQA, Winogrande, GPQA diamond, MATH level 5, OTIS Mock AIME 2024-2025, TriviaQA, and 13 more. The full table above shows each score with its evidence status.
4 of 19 recorded scores (21%) are independently reproduced rather than self-reported by the lab.
DeepSeek-V3 has 18 tracked claims: 3 independently reproduced, 10 self-reported, 4 unverified, 1 contradicted.
Listed API pricing: $0.27 per million input tokens, $1.1 per million output tokens. See the pricing block for the full breakdown.