DeepSeek-V3.2 vs Kimi K2 Instruct
Wins 6 of 7 areas
Coding · Reasoning · Facts · Math · Long documents · Following instructions
Wins 0 of 7 areas
—
DeepSeek-V3.2 is the stronger all-rounder.
Scores updated · 56 tests both models report · How we compare
Where each one wins
Tests won in each of the seven 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 software51DeepSeek-V3.25 of 6 tests
- ReasoningHard problems that need careful thinking30DeepSeek-V3.23 of 3 tests
- FactsGetting facts right instead of making them up21DeepSeek-V3.22 of 3 tests
- MathCompetition and research-level math10DeepSeek-V3.21 of 1 test
- Long documentsFinding answers in very long texts10DeepSeek-V3.21 of 1 test
- Following instructionsDoing exactly what it is asked10DeepSeek-V3.21 of 1 test
- AgentsCarrying out multi-step tasks on its own11Even1 each
Math, long documents 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.
DeepSeek-V3.2 costs 77% less for the same work.
The biggest differences
The three tests each model wins by the widest margin. Scores are out of 100.
Where DeepSeek-V3.2 pulls ahead
- Finds hard-to-locate facts by browsing the webBrowseComp+37.3points ahead
- Olympiad math problems with short answersIMOAnswerBench+32.5points ahead
- Recent programming contest problemsLiveCodeBench v6+29.6points ahead
Where Kimi K2 Instruct pulls ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+13points ahead
- Long, multi-file coding tasks in real codebasesSWE-bench Pro+12.1points ahead
- Real work tasks from 44 professionsGDPVal+8.6points ahead
Every test, side by side
All 56 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingDeepSeek-V3.2
- LiveCodeBench v6DeepSeek-V3.2 by 29.683.353.7+29.6
- SWE-bench MultilingualDeepSeek-V3.2 by 22.970.247.3+22.9
- SWE-bench ProKimi K2 Instruct by 12.115.627.7+12.1
- Terminal-Bench HardDeepSeek-V3.2 by 12.135.623.5+12.1
- SciCodeDeepSeek-V3.2 by 8.238.930.7+8.2
- SWE-bench VerifiedDeepSeek-V3.2 by 7.373.165.8+7.3
AgentsEven
- BrowseCompDeepSeek-V3.2 by 37.351.414.1+37.3
- GDPValKimi K2 Instruct by 8.69.618.2+8.6
ReasoningDeepSeek-V3.2
- Humanity's Last ExamDeepSeek-V3.2 by 18.224.66.4+18.2
- GPQA DiamondDeepSeek-V3.2 by 7.38476.7+7.3
- CritPtDeepSeek-V3.2 by 2.92.90+2.9
FactsDeepSeek-V3.2
- AA-Omniscience · Non-hallucinationKimi K2 Instruct by 1317.330.3+13
- Vectara HHEM hallucination ratelower is betterDeepSeek-V3.2 by 11.66.317.9+11.6
- AA-Omniscience · AccuracyDeepSeek-V3.2 by 7.63325.4+7.6
Following instructionsDeepSeek-V3.2
- IFBenchDeepSeek-V3.2 by 1960.741.7+19
Other results39 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.
- HMMT 2025DeepSeek-V3.2 by 51.490.238.8+51.4
- HMMT25 no toolsDeepSeek-V3.2 by 44.883.638.8+44.8
- AIME 2025DeepSeek-V3.2 by 43.693.149.5+43.6
- AIME25 no toolsDeepSeek-V3.2 by 38.389.351+38.3
- BrowseComp-ZHDeepSeek-V3.2 by 36.26528.8+36.2
- BrowseComp w/ toolsDeepSeek-V3.2 by 32.740.17.4+32.7
- LiveCodeBenchDeepSeek-V3.2 by 29.683.353.7+29.6
- BrowseComp-ZH w/ toolsDeepSeek-V3.2 by 25.747.922.2+25.7
- Frames w/ toolsDeepSeek-V3.2 by 22.180.258.1+22.1
- AA Agentic IndexKimi K2 Instruct by 19.418.337.7+19.4
- Artificial Analysis Coding IndexDeepSeek-V3.2 by 18.344.225.9+18.3
- LiveCodeBenchV6 no toolsDeepSeek-V3.2 by 1874.156.1+18
- τ²-Bench Telecom (AA run)DeepSeek-V3.2 by 17.290.673.4+17.2
- FinSearchComp-T3 (Tools-allowed)DeepSeek-V3.2 by 16.62710.4+16.6
- FinSearchComp-T3 w/ toolsDeepSeek-V3.2 by 16.62710.4+16.6
- τ²-BenchDeepSeek-V3.2 by 14.580.365.8+14.5
- HLE (with tools)DeepSeek-V3.2 by 13.940.826.9+13.9
- Seal-0 w/ toolsDeepSeek-V3.2 by 13.338.525.2+13.3
- OJ-Bench (cpp)DeepSeek-V3.2 by 12.738.225.5+12.7
- OJ-Bench (cpp) no toolsDeepSeek-V3.2 by 12.738.225.5+12.7
- vectara_factual_consistencyDeepSeek-V3.2 by 11.693.782.1+11.6
- Longform Writing eval (Kimi K2 Thinking system card)DeepSeek-V3.2 by 9.772.562.8+9.7
- Terminal-BenchDeepSeek-V3.2 by 7.737.730+7.7
- Terminal-Bench w/ simulated tools (JSON)Kimi K2 Instruct by 6.837.744.5+6.8
- AA IntelligenceDeepSeek-V3.2 by 6.221.515.3+6.2
- vectara_answer_rateKimi K2 Instruct by 692.698.6+6
- GPQA (unspecified)DeepSeek-V3.2 by 5.779.974.2+5.7
- xbench-DeepSearchKimi K2 Instruct by 5.355.761+5.3
- AA-OmniscienceDeepSeek-V3.2 by 4.1-22.5-26.6+4.1
- theagentcompanyDeepSeek-V3.2 by 43430+4
- MMLU-ProDeepSeek-V3.2 by 3.98581.1+3.9
- FinSearchComp-globalKimi K2 Instruct by 3.326.229.5+3.3
- GAIA (text only)DeepSeek-V3.2 by 3.363.560.2+3.3
- HealthBenchDeepSeek-V3.2 by 3.146.943.8+3.1
- HealthBench no toolsDeepSeek-V3.2 by 3.146.943.8+3.1
- Multi-SWE-BenchKimi K2 Instruct by 2.930.633.5+2.9
- vectara_avg_summary_lengthDeepSeek-V3.2 by 2.86259.2+2.8
- ArtifactsBenchDeepSeek-V3.2 by 1.655.854.2+1.6
- MMLU-ReduxDeepSeek-V3.2 by 193.792.7+1
Questions people ask
Which is better, DeepSeek-V3.2 or Kimi K2 Instruct?
DeepSeek-V3.2 wins six of the seven areas where both have results: coding, reasoning, facts, math, long documents and following instructions. Kimi K2 Instruct wins none. They are level on agents.
Which is better for coding?
DeepSeek-V3.2. It wins 5 of the 6 coding tests both models report; Kimi K2 Instruct wins 1.
Which is cheaper?
DeepSeek-V3.2 costs $0.28 per million input tokens and $0.42 per million output tokens; Kimi K2 Instruct costs $0.60 and $2.50. That makes DeepSeek-V3.2 about 77% cheaper for the same work.
How do you compare the two?
We use the 56 benchmark tests both models have published scores on. The verdict counts the 17 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 39 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.