DeepSeek-V3 vs Llama 3.1 Instruct 405B
Wins 3 of 6 areas
Coding · Long documents · Following instructions
Wins 0 of 6 areas
—
DeepSeek-V3 wins more areas, narrowly.
Scores updated · 89 tests both models report · How we compare
Where each one wins
Tests won in each of the six areas where both have results. Each piece is one test, so a longer bar means more evidence; grey means the two scored within a point of each other.
- CodingWriting and fixing software30DeepSeek-V33 of 3 tests
- Long documentsFinding answers in very long texts20DeepSeek-V32 of 2 tests
- Following instructionsDoing exactly what it is asked10DeepSeek-V31 of 1 test
- AgentsCarrying out multi-step tasks on its own00Even0 each · 1 tie
- ReasoningHard problems that need careful thinking11Even1 each · 2 ties
- FactsGetting facts right instead of making them up11Even1 each
Agents and following instructions rest on a single test each.
What it costs
Prices per million tokens, roughly 750,000 words. The bars show the cost of a million tokens read plus a million written.
The biggest differences
The tests each model wins by the widest margin, up to three each. Scores are out of 100.
Where DeepSeek-V3 pulls ahead
- Fixes real GitHub issues in Python projectsSWE-bench Verified+17.5points ahead
- Reasons across sets of long documentsAA-LCR+15.4points ahead
- Graduate-level biology, physics and chemistry questionsGPQA Diamond+14points ahead
Where Llama 3.1 Instruct 405B pulls ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+33.5points ahead
- Common-sense trick questionsSimpleBench+4.1points ahead
Every test, side by side
All 89 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingDeepSeek-V3
- SWE-bench VerifiedDeepSeek-V3 by 17.54224.5+17.5
- SciCodeDeepSeek-V3 by 9.13929.9+9.1
- Terminal-Bench HardDeepSeek-V3 by 8.415.26.8+8.4
ReasoningEven
- GPQA DiamondDeepSeek-V3 by 1465.551.5+14
- SimpleBenchLlama 3.1 Instruct 405B by 4.118.923+4.1
- Humanity's Last Examtie4.74tie
- CritPttie00tie
FactsEven
- AA-Omniscience · Non-hallucinationLlama 3.1 Instruct 405B by 33.514.147.6+33.5
- AA-Omniscience · AccuracyDeepSeek-V3 by 2.325.423.2+2.3
Long documentsDeepSeek-V3
- AA-LCRDeepSeek-V3 by 15.440.725.3+15.4
- LongBench v2DeepSeek-V3 by 12.648.736.1+12.6
Other results76 tests, not counted
Tests outside the eight areas. They are not counted above: several are summary scores built from other tests, or the same test under another name.
- CodeforcesDeepSeek-V3 by 1109 rating points113425.3+1109 rating
- τ²-Bench Telecom (AA run)DeepSeek-V3 by 28.147.119+28.1
- HumanEvalLlama 3.1 Instruct 405B by 23.865.289+23.8
- AA-OmniscienceLlama 3.1 Instruct 405B by 23.6-40.7-17.1+23.6
- Arena HardDeepSeek-V3 by 22.191.469.3+22.1
- LiveCodeBenchDeepSeek-V3 by 21.549.227.7+21.5
- AGIEvalDeepSeek-V3 by 1979.660.6+19
- AIR-Bench 2024Llama 3.1 Instruct 405B by 17.840.858.6+17.8
- Chinese SimpleQA (C-SimpleQA)DeepSeek-V3 by 17.66850.4+17.6
- MATHDeepSeek-V3 by 16.490.273.8+16.4
- MATH-500 (EM)DeepSeek-V3 by 16.490.273.8+16.4
- AIME 2024DeepSeek-V3 by 15.939.223.3+15.9
- Aider-Edit (Acc.)DeepSeek-V3 by 15.879.763.9+15.8
- CMMLUDeepSeek-V3 by 15.188.873.7+15.1
- CMMLU (Acc.)DeepSeek-V3 by 15.188.873.7+15.1
- C-EvalDeepSeek-V3 by 1486.572.5+14
- CCPMDeepSeek-V3 by 13.49278.6+13.4
- CMath (EM)DeepSeek-V3 by 13.490.777.3+13.4
- HarmBenchLlama 3.1 Instruct 405B by 1349.762.7+13
- MATH (EM)DeepSeek-V3 by 12.661.649+12.6
- MGSMLlama 3.1 Instruct 405B by 11.879.891.6+11.8
- MMLU-Pro (Acc.)DeepSeek-V3 by 11.664.452.8+11.6
- CRUXEval-O (output prediction)DeepSeek-V3 by 9.969.859.9+9.9
- MGSM (EM)DeepSeek-V3 by 9.979.869.9+9.9
- CRUXEval-I (input prediction)DeepSeek-V3 by 8.867.358.5+8.8
- Artificial Analysis Coding IndexDeepSeek-V3 by 8.52314.5+8.5
- CLUEWSCDeepSeek-V3 by 7.990.983+7.9
- SimpleQADeepSeek-V3 by 7.824.917.1+7.8
- RACE-MiddleLlama 3.1 Instruct 405B by 7.167.174.2+7.1
- MBPPDeepSeek-V3 by 775.468.4+7
- DROPDeepSeek-V3 by 6.891.684.8+6.8
- GSM8K (EM)DeepSeek-V3 by 5.889.383.5+5.8
- MBPP+ (EvalPlus-augmented)DeepSeek-V3 by 5.878.873+5.8
- MMMLUDeepSeek-V3 by 5.679.473.8+5.6
- MMMLU-non-English (Acc.)DeepSeek-V3 by 5.679.473.8+5.6
- RACE-HighLlama 3.1 Instruct 405B by 5.551.356.8+5.5
- HumanEval-Mul (Pass@1)DeepSeek-V3 by 5.482.677.2+5.4
- MMLU-Redux (Acc.)DeepSeek-V3 by 4.986.281.3+4.9
- NarrativeQADeepSeek-V3 by 4.779.674.9+4.7
- BBHDeepSeek-V3 by 4.687.582.9+4.6
- BBH (EM)DeepSeek-V3 by 4.687.582.9+4.6
- LiveCodeBench-Base (Pass@1)DeepSeek-V3 by 3.919.415.5+3.9
- simple_safety_testsLlama 3.1 Instruct 405B by 3.595.398.8+3.5
- FRAMES (Acc.)DeepSeek-V3 by 3.373.370+3.3
- DROP (F1)DeepSeek-V3 by 38986+3
- DROP (3-shot F1)DeepSeek-V3 by 2.991.688.7+2.9
- MMLU-ReduxDeepSeek-V3 by 2.989.186.2+2.9
- GSM8KLlama 3.1 Instruct 405B by 2.89496.8+2.8
- MMLU (Acc.)DeepSeek-V3 by 2.787.184.4+2.7
- MMLU-ProDeepSeek-V3 by 2.675.973.3+2.6
- IFEvalLlama 3.1 Instruct 405B by 2.586.188.6+2.5
- AA IntelligenceDeepSeek-V3 by 2.49.77.3+2.4
- bbqDeepSeek-V3 by 2.296.794.5+2.2
- ARC-ChallengeLlama 3.1 Instruct 405B by 1.695.396.9+1.6
- NaturalQuestionsLlama 3.1 Instruct 405B by 1.54041.5+1.5
- NaturalQuestions (EM)Llama 3.1 Instruct 405B by 1.54041.5+1.5
- OpenBookQADeepSeek-V3 by 1.495.494+1.4
- PIQALlama 3.1 Instruct 405B by 1.284.785.9+1.2
- PIQA (Acc.)Llama 3.1 Instruct 405B by 1.284.785.9+1.2
- XSTestDeepSeek-V3 by 1.297.195.9+1.2
- C3Llama 3.1 Instruct 405B by 1.178.679.7+1.1
- C3 (Acc.)Llama 3.1 Instruct 405B by 1.178.679.7+1.1
- naturalquestions_closedbookDeepSeek-V3 by 1.146.745.6+1.1
- IFEval (avg)tie87.386.4tie
- anthropic_red_teamtie97.196.5tie
- ARC-Easy (Acc.)tie98.998.4tie
- CMRCtie76.376tie
- HellaSwagtie88.989.2tie
- HellaSwag (Acc.)tie88.989.2tie
- WinoGrandetie84.985.2tie
- WinoGrande (Acc.)tie84.985.2tie
- TriviaQA (EM)tie82.982.7tie
- MMLUtie88.588.6tie
- ARC-Challenge (Acc.)tie95.395.3tie
- Pile-test (BPB)lower is bettertie0.50.5tie
- The Pile (Test, BPB)lower is bettertie0.50.5tie
Questions people ask
Which is better, DeepSeek-V3 or Llama 3.1 Instruct 405B?
DeepSeek-V3 wins three of the six areas where both have results: coding, long documents and following instructions. Llama 3.1 Instruct 405B wins none. They are level on agents, reasoning and facts.
Which is better for coding?
DeepSeek-V3. It wins 3 of the 3 coding tests both models report; Llama 3.1 Instruct 405B wins none.
How do you compare the two?
We use the 89 benchmark tests both models have published scores on. The verdict counts the 13 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 76 are listed but not counted, because several are summary scores or repeat a test. Each score is the one shown on the model's own page.