DeepSeek-V3 vs Kimi K2 Instruct
Wins 0 of 6 areas
—
Wins 5 of 6 areas
Coding · Agents · Reasoning · Long documents · Following instructions
Kimi K2 Instruct is the stronger all-rounder.DeepSeek-V3 is cheaper.
Scores updated · 51 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 software13Kimi K2 Instruct3 of 4 tests
- ReasoningHard problems that need careful thinking02Kimi K2 Instruct2 of 3 tests · 1 tie
- Following instructionsDoing exactly what it is asked01Kimi K2 Instruct1 of 2 tests · 1 tie
- AgentsCarrying out multi-step tasks on its own01Kimi K2 Instruct1 of 1 test
- Long documentsFinding answers in very long texts01Kimi K2 Instruct1 of 1 test
- FactsGetting facts right instead of making them up11Even1 each · 1 tie
Agents and long documents 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.
DeepSeek-V3 costs 63% 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 Kimi K2 Instruct pulls ahead
- Fixes real GitHub issues in Python projectsSWE-bench Verified+23.8points ahead
- Keeps track of context across a multi-turn chatMulti-Challenge+22.7points ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+16.2points ahead
Where DeepSeek-V3 pulls ahead
- Code for real scientific research problemsSciCode+8.3points ahead
Every test, side by side
All 51 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingKimi K2 Instruct
- SWE-bench VerifiedKimi K2 Instruct by 23.84265.8+23.8
- SciCodeDeepSeek-V3 by 8.33930.7+8.3
- Terminal-Bench HardKimi K2 Instruct by 8.315.223.5+8.3
- LiveCodeBench v6Kimi K2 Instruct by 6.846.953.7+6.8
ReasoningKimi K2 Instruct
- GPQA DiamondKimi K2 Instruct by 11.265.576.7+11.2
- Humanity's Last ExamKimi K2 Instruct by 1.74.76.4+1.7
- CritPttie00tie
FactsEven
- AA-Omniscience · Non-hallucinationKimi K2 Instruct by 16.214.130.3+16.2
- Vectara HHEM hallucination ratelower is betterDeepSeek-V3 by 11.86.117.9+11.8
- AA-Omniscience · Accuracytie25.425.4tie
Following instructionsKimi K2 Instruct
- Multi-ChallengeKimi K2 Instruct by 22.731.454.1+22.7
- IFBenchtie4141.7tie
Other results37 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.
- HarmBenchKimi K2 Instruct by 47.749.797.4+47.7
- AIR-Bench 2024Kimi K2 Instruct by 33.340.874.1+33.3
- τ²-BenchKimi K2 Instruct by 33.332.565.8+33.3
- CNMO 2024Kimi K2 Instruct by 31.143.274.3+31.1
- AIME 2024Kimi K2 Instruct by 30.439.269.6+30.4
- HumanEvalKimi K2 Instruct by 28.165.293.3+28.1
- τ²-Bench Telecom (AA run)Kimi K2 Instruct by 26.347.173.4+26.3
- vectara_avg_summary_lengthDeepSeek-V3 by 22.581.759.2+22.5
- Tau2 airlineKimi K2 Instruct by 17.53956.5+17.5
- AA-OmniscienceKimi K2 Instruct by 14.1-40.7-26.6+14.1
- vectara_factual_consistencyDeepSeek-V3 by 11.893.982.1+11.8
- HMMT 2025Kimi K2 Instruct by 11.327.538.8+11.3
- Aider-PolyglotKimi K2 Instruct by 10.449.660+10.4
- MATH-500 (EM)Kimi K2 Instruct by 7.290.297.4+7.2
- SimpleQAKimi K2 Instruct by 6.124.931+6.1
- AA IntelligenceKimi K2 Instruct by 5.69.715.3+5.6
- PolyMath-enKimi K2 Instruct by 5.659.565.1+5.6
- MMLU-ProKimi K2 Instruct by 5.275.981.1+5.2
- ZebraLogicKimi K2 Instruct by 58489+5
- simple_safety_testsKimi K2 Instruct by 4.795.3100+4.7
- LiveCodeBenchKimi K2 Instruct by 4.549.253.7+4.5
- LiveBenchKimi K2 Instruct by 472.476.4+4
- AceBenchKimi K2 Instruct by 3.872.776.5+3.8
- IFEvalKimi K2 Instruct by 3.786.189.8+3.7
- MMLU-ReduxKimi K2 Instruct by 3.689.192.7+3.6
- SuperGPQAKimi K2 Instruct by 3.553.757.2+3.5
- GSM8KKimi K2 Instruct by 3.39497.3+3.3
- OJBenchKimi K2 Instruct by 3.12427.1+3.1
- Artificial Analysis Coding IndexKimi K2 Instruct by 2.92325.9+2.9
- MultiPL-EKimi K2 Instruct by 2.683.185.7+2.6
- anthropic_red_teamKimi K2 Instruct by 2.297.199.3+2.2
- AIME 2025DeepSeek-V3 by 1.851.349.5+1.8
- τ²-Bench (Retail)Kimi K2 Instruct by 1.569.170.6+1.5
- vectara_answer_rateKimi K2 Instruct by 1.197.598.6+1.1
- XSTestKimi K2 Instruct by 1.197.198.2+1.1
- MMLUKimi K2 Instruct by 188.589.5+1
- AutoLogitie88.989.5tie
Questions people ask
Which is better, DeepSeek-V3 or Kimi K2 Instruct?
Kimi K2 Instruct wins five of the six areas where both have results: coding, agents, reasoning, long documents and following instructions. DeepSeek-V3 wins none, but costs 63% less. They are level on facts.
Which is better for coding?
Kimi K2 Instruct. It wins 3 of the 4 coding tests both models report; DeepSeek-V3 wins 1.
Which is cheaper?
DeepSeek-V3 costs $0.24 per million input tokens and $0.90 per million output tokens; Kimi K2 Instruct costs $0.60 and $2.50. That makes DeepSeek-V3 about 63% cheaper for the same work.
How do you compare the two?
We use the 51 benchmark tests both models have published scores on. The verdict counts the 14 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 37 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.