Atlas

Compare Models

DeepSeek V4.1 Flash vs Kimi K3.1 Pro

Kimi K3.1 Pro leads by 4.5 capability points (≈95% interval +2.3 to +6.6): a clear lead over DeepSeek V4.1 Flash · P(higher) ≈ 100%

Wins 81 of 100 shared benchmarks (17 lost, 2 tied)

Capability vs. relative cost
DeepSeek V4.1 FlashKimi K3.1 ProClosed weightsOpen weights

Each line runs through one model’s priced reasoning settings, cheapest to most expensive; grey points are other models’ priced settings in the same range. Results whose reasoning effort the source didn’t report are left out wherever a model has named effort levels. Capability and relative cost are the published estimates (relative cost is cost per task as a multiple of GPT-5 (high)); hover a setting, or use the arrow keys, to set it beside the other model’s same setting.

DeepSeek V4.1 FlashvsKimi K3.1 Procapability vs. relative costData Nov 15, 2026Village LLM Atlasatlas.village.dev

Benchmark wins17 – 81 · 100 shared, 2 tied

Comparing matching setups only (40 benchmarks excluded) · suites counted once
← DeepSeek V4.1 FlashKimi K3.1 Pro →

DeepSeek V4.1 FlashvsKimi K3.1 Prowins by categoryData Nov 15, 2026Data: Vals AI, Arena, Epoch AI and 24 othersVillage LLM Atlasatlas.village.dev

Margins

Each benchmark scaled 0–100% (its declared range, else the field’s range); Δ is Kimi K3.1 Pro minus DeepSeek V4.1 Flash.

IOI (Vals v2)Code40.381.9+41.6
FrontierMath Tier 4 (v2) - PrivateMath26.865.9+39.0
ProofBench v1.1Math54.091.0+37.0
Terminal-Bench 4.0Code31.265.7+34.5
Creative Writing v3 EloChat & Writing15402173+32.0
Terminal-Bench Science 0.1 (Vals)Science0.031.4+31.4
Concept Synth INDUCTION Challenge64 Round 1General QA17.246.9+29.7
BuseyBench SVGMultimodal4.16.9+28.5
τ³-Banking (Fireworks, banking_knowledge)Agents35.763.3+27.6
SRE Bench — Fully SolvedCode0.827.1+26.3
APEX-Agents 1.1Professional Work50.575.9+25.4
VoxelBench Text RatingGames18612216+25.3
EnterpriseBench: CoreCraftProfessional Work41.566.1+24.6
ArXivMath 09/2026 (Native Harnesses)Math40.664.2+23.6
Concept Synth INDUCTION Challenge100 Round 1General QA11.034.0+23.0
τ³-Banking (Mercor run, bm25_grep retrieval)Agents37.859.8+22.0
Furniture Assembly BenchmarkMultimodal34.255.8+21.7
ArXivMath 08/2026 (Native Harnesses)Math43.964.9+21.0
Humanity's Last Exam DiamondGeneral QA19.139.4+20.3
SimpleQA Verified (Epoch anti-abstention)General QA42.962.2+19.3
Legal Research Bench Civil LitigationLegal48.129.6−18.5
Legal Research Bench ImmigrationLegal36.418.2−18.2
LMArena Agent Arena Work Net Improvement (Steering Burden)Agents2.59.6+17.2
TCSExamMath40.457.4+17.0
GameBench 2Games68.484.0+15.6
LMArena Agent Arena Code Net Improvement (Steering Burden)Agents2.79.3+14.9
LMArena Agent Arena Chat Net Improvement (Steering Burden)Agents1.67.7+14.8
Vals Science IndexIndexes29.644.4+14.7
AA-Briefcase v1.1 EloAgents14151678+14.4
Humanity's Last ExamGeneral QA36.851.1+14.3
Bug Hunt BenchCode20.634.3+13.7
Legal Research Bench Weighted PassLegal84.571.4−13.1
Conceptual Reasoning IndexIndexes58.467.8+12.9
Creative Writing v3 Rubric ScoreChat & Writing14.116.6+12.9
ORCA-bench Hard RCA (Fireworks 324-task subset)Agents20.332.9+12.7
APEX-Accounting 1.1Finance & Accounting50.162.4+12.3
Legal Research Bench Business/CommercialLegal39.527.9−11.6
ArXivLean 09/2026Math5.717.0+11.3
ProgramBench - almost solved rollout percentageCode20.331.5+11.2
FrontierMath Tiers 1-3 (v2) - PrivateMath67.478.2+10.9
GDPval-AA v2.1Professional Work16001846+10.8
ALE-BenchCode18292132+10.7
LMCAGeneral QA47.057.6+10.6
DAXBench v2Code75.586.2+10.6
LMArena Agent Arena Work Praise vs ComplaintAgents3.724.6+10.5
LMArena Agent Arena Chat Praise vs ComplaintAgents3.123.9+10.4
MineBench — Bradley-Terry RatingGames18221992+10.3
GBENCH One-shot Coding GScoreGames48.058.2+10.2
Harvey LAB held-out set (Fireworks run)Legal5.715.4+9.7
BrokenArXiv 08/2026 (Native Harnesses)Math54.864.3+9.5
ACCoRDGeneral QA68.177.5+9.4
BrokenArXiv 09/2026 (Native Harnesses)Math54.763.5+8.8
LMArena Code Arena WebDevCode16211688+8.8
LMArena Agent Arena Code Praise vs ComplaintAgents3.019.0+8.0
BioMysteryBench (Vals Terminus 2)Science67.875.6+7.8
BigFinanceBench private holdout (final-answer accuracy)Finance & Accounting40.648.2+7.6
LMArena Agent Arena Code Steering BurdenAgents2.314.9+6.3
Design Arena Overall Code CategoriesCode13161353+6.1
Longform Creative WritingChat & Writing74.079.9+5.9
ExploitGym - release-reported percentage, cohort unspecifiedCybersecurity15.320.7+5.4
Legacy-Bench (Fireworks, avg@3)Code57.763.0+5.3
Legal Research Bench Constitutional/Civil RightsLegal36.831.6−5.3
Legal Research Bench CriminalLegal33.328.1−5.3
LMArena Agent Arena Chat Bash ReliabilityAgents-1.09.2+5.1
LMArena Agent Arena Work Steering BurdenAgents1.511.5+5.0
MysteryMechanismScience21.226.1+5.0
Epoch Capabilities Index (ECI)Indexes156162+4.5
AutomationBench-AA 1.0.6Agents68.973.3+4.4
Finance Agent v2.1Finance & Accounting65.461.5−3.9
LMArena Agent Arena Work Bash ReliabilityAgents0.17.6+3.7
LMArena Agent Arena Code Bash ReliabilityAgents-0.26.8+3.5
DTBenchGeneral QA89.993.3+3.5
Access to Justice BenchmarkLegal72.969.5−3.4
Epiq LAER CorporateBenchProfessional Work75.378.4+3.2
Legal Research Bench Administrative/RegulatoryLegal53.056.1+3.0
Chess Puzzles (task version 2.0)Games48.045.0−3.0
DeepSWE v1.1Code72.675.5+2.9
CloudOpsBench v1.0Agents21.819.1−2.7
Harvey LAB: In-House Contracting (held-out)Legal0.33.0+2.7
LMArena Agent Arena Chat Steering BurdenAgents-0.54.5+2.5
SPU (Science Prediction Universe)Science0.20.2−2.4
Excel Modeling Benchmark v1.1Finance & Accounting67.870.0+2.2
Political Science LLM BenchmarkProfessional Work66.568.6+2.1
LMArena Agent Arena Chat Confirmed SuccessAgents4.91.1−1.9
DualEntry Accounting AI BenchmarkFinance & Accounting84.286.1+1.9
OTIS Mock AIME 2024-2025Math98.3100.0+1.7
LMArena Agent Arena Work Confirmed SuccessAgents5.72.8−1.4
Tax Agent BenchFinance & Accounting62.563.9+1.4
LMArena Text Arena EloChat & Writing14751483+1.4
LMArena Agent Arena Code Confirmed SuccessAgents6.63.8−1.4
LiveBench 2.0Agents78.076.7−1.3
OpenWorldEngineer (Mean@8)Code12.413.6+1.2
Vals Finance IndexIndexes57.156.0−1.1
Mystery Game PuzzlesGames43.042.0−1.0
Vibe Code Bench v1.1Code84.784.9+0.2
LMArena Agent Arena Work Tool HallucinationAgents0.50.7+0.1
LMArena Agent Arena Code Tool HallucinationAgents0.50.7+0.1
LMArena Agent Arena Chat Tool HallucinationAgents0.30.5+0.1
Legal Research Bench FamilyLegal13.613.60.0
Legal Research Bench HealthLegal73.373.30.0

DeepSeek V4.1 FlashvsKimi K3.1 Pronormalised score marginsData Nov 15, 2026Data: Vals AI, Arena, Epoch AI and 24 othersVillage LLM Atlasatlas.village.dev

Win counts use each benchmark’s declared score direction. Represented suite slices and the derived Capability score are excluded from headline totals. Missing or incompatible setups are excluded, never counted as losses.
Loading Atlas data…