// COMPARISON

DeepSeek-V3.1-Terminus vs Kimi K2 Thinking

A head-to-head benchmark comparison of DeepSeek-V3.1-Terminus vs Kimi K2 Thinking across 49 evaluations. Kimi K2 Thinking leads on the most benchmarks (26).

The models

Leads 26
MO

Kimi K2 Thinking

Moonshot
Language Open (Restricted)
Params
1T
Context
β€”
Released
Nov 2025

Today, we are introducing Kimi K2 Thinking, our best open-source thinking model. Built as a thinking agent, it reasons step by step while using tools, achieving.

Full Kimi K2 Thinking specs β†’

Benchmark comparison

Benchmark DeepSeek-V3.1-Terminus Kimi K2 Thinking
agentic
BrowseComp β€” 60.2%
BrowseComp-Zh β€” 62.3
FinSearchComp-T3 β€” 47.4
MCP-Mark β€” 20.4
MCP-Universe β€” 35.6
Seal-0 β€” 56.3
Terminal-Bench 2.0 β€” 47.1%
Tool-Decathlon β€” 17.6
τ²-Bench (TAU2) β€” 74.3
code
Codeforces Elo 2,073 β€”
FullStackBench 55.48% β€”
LiveCodeBench 48.02% 82.6%
LiveCodeBench v6 75.33% 83.1%
MBPP 90.69% β€”
SWE-bench Verified β€” 71.3%
SciCode β€” 44.8%
coding
Aider Benchmark 92.86 β€”
Multi-SWE-bench β€” 41.9
MultiPL-E 77.68 β€”
OJ-Bench (cpp) β€” 48.7
SWE-Bench Multilingual β€” 61.1
general
ArenaHard v2 63.24 β€”
C-Eval 91.76% β€”
IFEval 89.09% β€”
Longform Writing β€” 73.8
MMLU-Pro 85% 84.6%
MMLU-Redux 92.37% 94.4%
MultiChallenge 45.79% β€”
knowledge
AGIEval 89.83 β€”
CMMLU 89.2 β€”
TriviaQA 82.77 β€”
long-context
Frames β€” 87
math
AIME 2024 71.67% β€”
AIME 2025 89.06% 100%
CNMO 2024 85.42% β€”
HMMT February 2025 86.1% 97.5%
HMMT November 2025 β€” 89.2
IMOAnswerBench β€” 78.6
Omni-MATH 81.93 β€”
UGMathBench 77.19 β€”
reasoning
ARC-AGI-1 40.62 β€”
BBEH 61.04 β€”
FinanceReasoning 87.76 β€”
GPQA Diamond 81% 84.5%
Humanity's Last Exam 17.82% 51%
ZebraLogic 96.33% β€”
safety
HealthBench 50.19% 58%
science
PhyBench 47.91 β€”
tool-use
BFCL v3 62.01% β€”

Best result per row highlighted in cyan. Each benchmark links to its definition and sources; each model links to its full scorecard.

Frequently asked questions

Which model is best, DeepSeek-V3.1-Terminus vs Kimi K2 Thinking? β–Ά

Across the 49 benchmarks compared here, Kimi K2 Thinking leads on the most (26). The right choice still depends on your task β€” check the per-benchmark table above for coding, reasoning and math.

Which of these models is best for coding? β–Ά

DeepSeek-V3.1-Terminus has the strongest result on Codeforces Elo among the models compared here. See the code-category rows in the table for the full picture.

How many benchmarks are compared? β–Ά

This comparison covers 49 benchmarks on which at least one of the selected models has a published score.

Where do the benchmark scores come from? β–Ά

Scores are aggregated from official model cards, technical reports and standard public evaluations, and link back to each benchmark's source. They are updated as new results are published.

Explore further

β†’ DeepSeek-V3.1-Terminus β†’ Kimi K2 Thinking β†’ AI benchmarks leaderboard β†’ Build a different comparison
Build a different comparison

Models (0 selected)

Benchmarks (auto)

Select at least two models.