DeepSeek-V3.2-Speciale vs Kimi K2 Thinking
Wins 2 of 6 areas
Coding · Reasoning
Wins 3 of 6 areas
Agents · Long documents · Following instructions
Kimi K2 Thinking wins more areas, narrowly.DeepSeek-V3.2-Speciale is cheaper and better at reasoning.
Scores updated · 22 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.
- AgentsCarrying out multi-step tasks on its own01Kimi K2 Thinking1 of 1 test
- Long documentsFinding answers in very long texts01Kimi K2 Thinking1 of 1 test
- Following instructionsDoing exactly what it is asked01Kimi K2 Thinking1 of 1 test
- ReasoningHard problems that need careful thinking40DeepSeek-V3.2-Speciale4 of 4 tests
- CodingWriting and fixing software30DeepSeek-V3.2-Speciale3 of 3 tests
- FactsGetting facts right instead of making them up11Even1 each
Agents, 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-Speciale 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 Kimi K2 Thinking pulls ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+13.6points ahead
- Follows unfamiliar, precisely checkable instructionsIFBench+4.2points ahead
- Reasons across sets of long documentsAA-LCR+2points ahead
Where DeepSeek-V3.2-Speciale pulls ahead
- Common-sense trick questionsSimpleBench+26.3points ahead
- Fixes real GitHub issues in Python projectsSWE-bench Verified+9.7points ahead
- Answers hard knowledge questions correctlyAA-Omniscience · Accuracy+6.9points ahead
Every test, side by side
All 22 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingDeepSeek-V3.2-Speciale
- SWE-bench VerifiedDeepSeek-V3.2-Speciale by 9.773.163.4+9.7
- Terminal-Bench HardDeepSeek-V3.2-Speciale by 3.734.831.1+3.7
- SciCodeDeepSeek-V3.2-Speciale by 1.64442.4+1.6
ReasoningDeepSeek-V3.2-Speciale
- SimpleBenchDeepSeek-V3.2-Speciale by 26.352.626.3+26.3
- Humanity's Last ExamDeepSeek-V3.2-Speciale by 4.928.723.8+4.9
- CritPtDeepSeek-V3.2-Speciale by 4.87.42.6+4.8
- GPQA DiamondDeepSeek-V3.2-Speciale by 3.387.183.8+3.3
FactsEven
- AA-Omniscience · Non-hallucinationKimi K2 Thinking by 13.610.824.4+13.6
- AA-Omniscience · AccuracyDeepSeek-V3.2-Speciale by 6.937.830.9+6.9
Following instructionsKimi K2 Thinking
- IFBenchKimi K2 Thinking by 4.263.968.1+4.2
Other results10 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.
- τ²-Bench Telecom (AA run)Kimi K2 Thinking by 93093+93
- AA Agentic IndexKimi K2 Thinking by 47.9047.9+47.9
- Terminal-Bench 2.0DeepSeek-V3.2-Speciale by 10.746.435.7+10.7
- HMMT 2025DeepSeek-V3.2-Speciale by 9.899.289.4+9.8
- AA IntelligenceKimi K2 Thinking by 7.514.522+7.5
- HMMT Feb. 2025DeepSeek-V3.2-Speciale by 4.297.593.3+4.2
- HMMT Nov. 2025DeepSeek-V3.2-Speciale by 4.193.389.2+4.1
- AA-OmniscienceDeepSeek-V3.2-Speciale by 3.7-17.7-21.4+3.7
- Artificial Analysis Coding IndexDeepSeek-V3.2-Speciale by 3.137.934.8+3.1
- AIME 2025DeepSeek-V3.2-Speciale by 1.59694.5+1.5
Questions people ask
Which is better, DeepSeek-V3.2-Speciale or Kimi K2 Thinking?
Kimi K2 Thinking wins three of the six areas where both have results: agents, long documents and following instructions. DeepSeek-V3.2-Speciale wins coding and reasoning, and costs 77% less. They are level on facts.
Which is better for coding?
DeepSeek-V3.2-Speciale. It wins 3 of the 3 coding tests both models report; Kimi K2 Thinking wins none.
Which is cheaper?
DeepSeek-V3.2-Speciale costs $0.29 per million input tokens and $0.43 per million output tokens; Kimi K2 Thinking costs $0.60 and $2.50. That makes DeepSeek-V3.2-Speciale about 77% cheaper for the same work.
How do you compare the two?
We use the 22 benchmark tests both models have published scores on. The verdict counts the 12 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 10 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.