DeepSeek-V3 vs Qwen2.5 Instruct 72B
Wins 5 of 5 areas
Coding · Reasoning · Facts · Long documents · Following instructions
Wins 0 of 5 areas
—
DeepSeek-V3 is the stronger all-rounder.Qwen2.5 Instruct 72B is cheaper.
Scores updated · 90 tests both models report · How we compare
Where each one wins
Tests won in each of the five 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
- ReasoningHard problems that need careful thinking20DeepSeek-V32 of 3 tests · 1 tie
- Long documentsFinding answers in very long texts20DeepSeek-V32 of 2 tests
- FactsGetting facts right instead of making them up10DeepSeek-V31 of 2 tests · 1 tie
- Following instructionsDoing exactly what it is asked10DeepSeek-V31 of 1 test
Following instructions rests on a single test.
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.
Qwen2.5 Instruct 72B costs 15% less for the same work.
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
- Reasons across sets of long documentsAA-LCR+20.7points ahead
- Fixes real GitHub issues in Python projectsSWE-bench Verified+18.2points ahead
- Graduate-level biology, physics and chemistry questionsGPQA Diamond+16.4points ahead
Where Qwen2.5 Instruct 72B pulls ahead
No clear win on a test scored out of 100.
Every test, side by side
All 90 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 18.24223.8+18.2
- SciCodeDeepSeek-V3 by 12.33926.7+12.3
- Terminal-Bench HardDeepSeek-V3 by 10.715.24.5+10.7
ReasoningDeepSeek-V3
- GPQA DiamondDeepSeek-V3 by 16.465.549.1+16.4
- Humanity's Last ExamDeepSeek-V3 by 1.14.73.6+1.1
- CritPttie00tie
FactsDeepSeek-V3
- AA-Omniscience · AccuracyDeepSeek-V3 by 825.417.5+8
- AA-Omniscience · Non-hallucinationtie14.114.5tie
Long documentsDeepSeek-V3
- AA-LCRDeepSeek-V3 by 20.740.720+20.7
- LongBench v2DeepSeek-V3 by 9.348.739.4+9.3
Following instructionsDeepSeek-V3
- IFBenchDeepSeek-V3 by 4.14136.9+4.1
Other results79 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 points113424.8+1109 rating
- HarmBenchQwen2.5 Instruct 72B by 23.149.772.8+23.1
- HumanEvalQwen2.5 Instruct 72B by 21.465.286.6+21.4
- LiveBenchDeepSeek-V3 by 20.172.452.3+20.1
- Chinese SimpleQA (C-SimpleQA)DeepSeek-V3 by 19.66848.4+19.6
- AIR-Bench 2024Qwen2.5 Instruct 72B by 18.240.859+18.2
- AIME 2024DeepSeek-V3 by 15.939.223.3+15.9
- SimpleQADeepSeek-V3 by 15.824.99.1+15.8
- DROPDeepSeek-V3 by 14.991.676.7+14.9
- DROP (3-shot F1)DeepSeek-V3 by 14.991.676.7+14.9
- Aider-Edit (Acc.)DeepSeek-V3 by 14.379.765.4+14.3
- MBPPQwen2.5 Instruct 72B by 12.875.488.2+12.8
- τ²-Bench Telecom (AA run)DeepSeek-V3 by 12.647.134.5+12.6
- AA-OmniscienceDeepSeek-V3 by 12.4-40.7-53.1+12.4
- MMLUDeepSeek-V3 by 11.588.577+11.5
- Artificial Analysis Coding IndexDeepSeek-V3 by 11.12311.9+11.1
- TriviaQA (EM)DeepSeek-V3 by 1182.971.9+11
- naturalquestions_closedbookDeepSeek-V3 by 10.846.735.9+10.8
- Arena HardDeepSeek-V3 by 10.291.481.2+10.2
- MATH-500 (EM)DeepSeek-V3 by 10.290.280+10.2
- CRUXEval-O (output prediction)DeepSeek-V3 by 9.969.859.9+9.9
- CLUEWSCDeepSeek-V3 by 8.490.982.5+8.4
- DROP (F1)DeepSeek-V3 by 8.48980.6+8.4
- CRUXEval-I (input prediction)DeepSeek-V3 by 8.267.359.1+8.2
- MultiPL-EDeepSeek-V3 by 883.175.1+8
- BBHDeepSeek-V3 by 7.787.579.8+7.7
- BBH (EM)DeepSeek-V3 by 7.787.579.8+7.7
- MATH (EM)DeepSeek-V3 by 7.261.654.4+7.2
- NaturalQuestionsDeepSeek-V3 by 6.84033.2+6.8
- NaturalQuestions (EM)DeepSeek-V3 by 6.84033.2+6.8
- LongBench v2 overall (w/o CoT)DeepSeek-V3 by 6.648.742.1+6.6
- LiveCodeBench-Base (Pass@1)DeepSeek-V3 by 6.519.412.9+6.5
- LiveCodeBenchQwen2.5 Instruct 72B by 6.349.255.5+6.3
- CMath (EM)DeepSeek-V3 by 6.290.784.5+6.2
- MMLU-Pro (Acc.)DeepSeek-V3 by 6.164.458.3+6.1
- HumanEval-Mul (Pass@1)DeepSeek-V3 by 5.382.677.3+5.3
- NarrativeQADeepSeek-V3 by 5.179.674.5+5.1
- MMLU-ProDeepSeek-V3 by 4.875.971.1+4.8
- simple_safety_testsQwen2.5 Instruct 72B by 4.795.3100+4.7
- MMMLUDeepSeek-V3 by 4.679.474.8+4.6
- MMMLU-non-English (Acc.)DeepSeek-V3 by 4.679.474.8+4.6
- HellaSwagDeepSeek-V3 by 4.188.984.8+4.1
- HellaSwag (Acc.)DeepSeek-V3 by 4.188.984.8+4.1
- AGIEvalDeepSeek-V3 by 3.879.675.8+3.8
- MGSMDeepSeek-V3 by 3.679.876.2+3.6
- MGSM (EM)DeepSeek-V3 by 3.679.876.2+3.6
- CCPMDeepSeek-V3 by 3.59288.5+3.5
- FRAMES (Acc.)DeepSeek-V3 by 3.573.369.8+3.5
- MMLU-Redux (Acc.)DeepSeek-V3 by 386.283.2+3
- C-EvalQwen2.5 Instruct 72B by 2.786.589.2+2.7
- WinoGrandeDeepSeek-V3 by 2.684.982.3+2.6
- WinoGrande (Acc.)DeepSeek-V3 by 2.684.982.3+2.6
- anthropic_red_teamQwen2.5 Instruct 72B by 2.597.199.6+2.5
- MMLU-ReduxDeepSeek-V3 by 2.389.186.8+2.3
- MMLU (Acc.)DeepSeek-V3 by 2.187.185+2.1
- PIQADeepSeek-V3 by 2.184.782.6+2.1
- PIQA (Acc.)DeepSeek-V3 by 2.184.782.6+2.1
- AA IntelligenceDeepSeek-V3 by 29.77.7+2
- IFEvalDeepSeek-V3 by 286.184.1+2
- C3DeepSeek-V3 by 1.978.676.7+1.9
- C3 (Acc.)DeepSeek-V3 by 1.978.676.7+1.9
- GSM8KQwen2.5 Instruct 72B by 1.89495.8+1.8
- MATHDeepSeek-V3 by 1.890.288.4+1.8
- MBPP+ (EvalPlus-augmented)DeepSeek-V3 by 1.878.877+1.8
- bbqDeepSeek-V3 by 1.396.795.4+1.3
- GSM8K (EM)DeepSeek-V3 by 189.388.3+1
- RACE-HighDeepSeek-V3 by 151.350.3+1
- RACE-MiddleQwen2.5 Instruct 72B by 167.168.1+1
- ARC-Challengetie95.394.5tie
- ARC-Challenge (Acc.)tie95.394.5tie
- OpenBookQAtie95.496.2tie
- XSTesttie97.197.9tie
- CMMLUtie88.889.5tie
- CMMLU (Acc.)tie88.889.5tie
- ARC-Easy (Acc.)tie98.998.4tie
- CMRCtie76.375.8tie
- IFEval (avg)tie87.387.2tie
- Pile-test (BPB)lower is bettertie0.50.6tie
- The Pile (Test, BPB)lower is bettertie0.50.6tie
Questions people ask
Which is better, DeepSeek-V3 or Qwen2.5 Instruct 72B?
DeepSeek-V3 wins all five areas where both have results: coding, reasoning, facts, long documents and following instructions. Qwen2.5 Instruct 72B wins none, but costs 15% less.
Which is better for coding?
DeepSeek-V3. It wins 3 of the 3 coding tests both models report; Qwen2.5 Instruct 72B wins none.
Which is cheaper?
DeepSeek-V3 costs $0.24 per million input tokens and $0.90 per million output tokens; Qwen2.5 Instruct 72B costs $0.47 and $0.49. That makes Qwen2.5 Instruct 72B about 15% cheaper for the same work.
How do you compare the two?
We use the 90 benchmark tests both models have published scores on. The verdict counts the 11 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 79 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.