DeepSeek V4.1 Flash 1017 vs Kimi K3
Full profiles: DeepSeek V4.1 Flash 1017 (DeepSeek) and Kimi K3 (Moonshot AI): every benchmark, price and rank.
Updated · Data from 9 sources
As of 24 Oct 2026, DeepSeek V4.1 Flash 1017 (max) and Kimi K3 (max) are neck and neck on quality (64.0 vs 64.7 on the BenchLeader Index, inside the ±5.1 margin of error), and DeepSeek V4.1 Flash 1017 (max) costs 11× less, which makes it the better buy for most workloads. Pick DeepSeek V4.1 Flash 1017 for coding (70 vs 65); pick Kimi K3 for knowledge (63 vs 57).
DeepSeek V4.1 Flash 1017 benchmarks max DeepSeek
Kimi K3 benchmarks max Moonshot AIOpen in the comparison toolWhich to pick for…
The pick's numbers come first. Category scores on the BenchLeader scale (50 is the average model, 15 points one standard deviation). High confidence means a lead of half a standard deviation or more; under one point is called either.
| Use case | Pick | Why | Confidence |
|---|---|---|---|
| Coding | DeepSeek V4.1 Flash 1017 | 70 vs 65 category score; Arena WebDev 1664 vs 1655 | Medium |
| Agents & tool use | DeepSeek V4.1 Flash 1017 | 63 vs 59 category score; CWE-Bench (AA) 69.2% vs 53.3% | Medium |
| Maths | Kimi K3 | 63 vs 62 category score; FrontierMath Tiers 1–3 72.2% vs 71.2% | Medium |
| Reasoning | DeepSeek V4.1 Flash 1017 | 71 vs 66 category score; GPQA Diamond 93.2% vs 93.1% | Medium |
| Knowledge | Kimi K3 | 63 vs 57 category score; SimpleQA Verified 50.6% vs 42.5% | Medium |
| Long documents | Kimi K3 | 70 vs 66 category score; AA-LCR 88.7% vs 82.7% | Medium |
| Images and other media | Either | 64 vs 64 category score; MMMU-Pro 80.6% vs 80.5% | Either |
| Chat quality (human preference) | – | Only Kimi K3 has an Arena rating (68 category score) | Not comparable |
| Lowest cost | DeepSeek V4.1 Flash 1017 | $0.525 vs $6.00 per 1M blended; agentic coding session $0.023 vs $0.240 | – |
| Fastest first answer | DeepSeek V4.1 Flash 1017 | 11.2 s (DeepSeek V4.1 Flash 1017 (max)) vs 55.4 s (Kimi K3 (max)) to the first answer token | – |
| High volume on a budget | Neither flagship | A smaller model is usually the better fit: Cheapest good LLMs, Fastest LLMs | – |
Key differences
- DeepSeek V4.1 Flash 1017 (max) and Kimi K3 (max) are neck and neck on quality (64.0 vs 64.7, within the ±5.1 margin).
- DeepSeek V4.1 Flash 1017 (max) is stronger in agents & tools, coding, composite, multimodal, reasoning.
- Kimi K3 (max) is stronger in knowledge, long context, maths, human preference.
- DeepSeek V4.1 Flash 1017 (max) is 11× cheaper ($0.525 vs $6.00 per 1M blended).
- DeepSeek V4.1 Flash 1017 (max) streams 5.2× faster (198 vs 39 tokens per second).
At a glance
| Spec | DeepSeek V4.1 Flash 1017 | Kimi K3 |
|---|---|---|
| Developer | DeepSeek | Moonshot AI |
| Released | 17 Oct 2026 | 16 Jul 2026 |
| Context window (tokens) | 1,048,576 | 1,048,576 |
| Compared at | DeepSeek V4.1 Flash 1017 (max) | Kimi K3 (max) |
| Open weights | Yes | Yes |
| Price in / out / cached input per 1M | $0.300 / $1.20 / $0.0060 | $3.00 / $15.00 / $0.300 |
Category by category
Category scores on the BenchLeader scale: 50 is the average evaluated model, 15 points is one standard deviation. A gap under one point is called level.
| Category | DeepSeek V4.1 Flash 1017 (max) | Kimi K3 (max) | Leader |
|---|---|---|---|
| Reasoning | 70.6 | 66.4 | DeepSeek V4.1 Flash 1017 |
| Coding | 70.1 | 64.6 | DeepSeek V4.1 Flash 1017 |
| Agents & tools | 63.0 | 58.9 | DeepSeek V4.1 Flash 1017 |
| Maths | 61.8 | 63.4 | Kimi K3 |
| Knowledge | 57.1 | 62.8 | Kimi K3 |
| Human preference | – | 68.1 | not comparable |
| Multimodal | 63.7 | 63.6 | level |
| Long context | 66.4 | 69.6 | Kimi K3 |
| Composite | 80.0 | 78.3 | DeepSeek V4.1 Flash 1017 |
Where sources disagree
Across the 43 benchmarks both have, DeepSeek V4.1 Flash 1017 leads 29, Kimi K3 leads 14. Every overall index that rates both puts DeepSeek V4.1 Flash 1017 ahead.
| Overall index | DeepSeek V4.1 Flash 1017 (max) | Kimi K3 (max) | Leader |
|---|---|---|---|
| BenchLeader Index | 64.0 | 64.7 | neck and neck |
| AA Intelligence Index v4.3.2 | 45.1 | 43.6 | DeepSeek V4.1 Flash 1017 |
| Vals Index | 61.3 | 53.0 | DeepSeek V4.1 Flash 1017 |
Each publisher runs its own tasks, harnesses and effort settings, so their overall verdicts can differ; BenchLeader reconciles them on one scale. † measured at another effort setting of the same model (the publisher did not test the one compared here).
Price, speed and latency
List prices blended at three input tokens per output token. Per-task prices use the token counts on the methodology page; suite cost is what Artificial Analysis paid to run its whole Intelligence Index.
| Measure | DeepSeek V4.1 Flash 1017 (max) | Kimi K3 (max) |
|---|---|---|
| Blended price per 1M tokens | $0.525 | $6.00 |
| Input / output per 1M | $0.300 / $1.20 | $3.00 / $15.00 |
| Cached input per 1M | $0.0060 | $0.300 |
| Price per task: chat reply | $0.0005 | $0.0057 |
| Price per task: summarize a 30-page report | $0.0043 | $0.045 |
| Price per task: code edit | $0.0036 | $0.041 |
| Price per task: agentic coding session | $0.023 | $0.240 |
| Price per task: structured extraction | $0.0008 | $0.0090 |
| Cost to run a full benchmark suite | $0.262 | $2.00 |
| Output speed | 198 tok/s | 39 tok/s |
| Time to first answer | 11.2 s | 55.4 s |
| Time to a complete answer | 13.7 s | 68.4 s |
Per token is not per task: on Artificial Analysis' full suite DeepSeek V4.1 Flash 1017 (max) cost $0.262 and Kimi K3 (max) $2.00, because DeepSeek V4.1 Flash 1017 (max) used about 1.5× as many tokens for the same work, which a per-token price hides.
Each model is compared at its best-ranked reasoning-effort setting. DeepSeek V4.1 Flash 1017 (max) scores 64.0 and takes 11 s to its first answer token. Kimi K3 (max) scores 64.7 and takes 55 s to its first answer token; Kimi K3 scores 59.1 in 1 s, Kimi K3 (low) scores 58.2 in 57 s. Higher effort means more thinking before the answer: better scores, longer waits and more tokens billed. How reasoning effort changes scores and cost.
Every benchmark
133 benchmarks, grouped by category; each name links to the board and its leaders. † measured at another effort setting of the same model, shown for reference and not counted as a win.
| Metric | Publisher | DeepSeek V4.1 Flash 1017 (max) | Kimi K3 (max) |
|---|---|---|---|
| BenchLeader Index | BenchLeader | 64.0 | 64.7 |
| Agents & tools score | BenchLeader | 63.0 | 58.9 |
| Coding score | BenchLeader | 70.1 | 64.6 |
| Composite score | BenchLeader | 80.0 | 78.3 |
| Knowledge score | BenchLeader | 57.1 | 62.8 |
| Long context score | BenchLeader | 66.4 | 69.6 |
| Maths score | BenchLeader | 61.8 | 63.4 |
| Multimodal score | BenchLeader | 63.7 | 63.6 |
| Reasoning score | BenchLeader | 70.6 | 66.4 |
| Human preference score | BenchLeader | – | 68.1 |
| Blended price $/M | BenchLeader | $0.525 | $6.00 |
| Output speed | BenchLeader | 198 tok/s | 39 tok/s |
| Time to first answer | BenchLeader | 11.2 s | 55.4 s |
| Context window | BenchLeader | 1,048,576 tokens | 1,048,576 tokens |
| Reasoning | |||
| GPQA Diamond | Epoch AI Benchmarking Hub | 93.2% | 93.1% |
| SimpleBench | SimpleBench | – | 60.7% |
| Arena Hard Prompts | Arena (LMArena) | – | 1516 |
| LiveBench 2.0 Reasoning | LiveBench | 89.1%† | 90.7%† |
| LiveBench 1.0 Reasoning | LiveBench | – | 90.7%† |
| GPQA Diamond (AA) | Artificial Analysis | – | 93.5% |
| Humanity's Last Exam (AA) | Artificial Analysis | 43.3% | 46.9% |
| GPQA Diamond (Vals) | Vals AI | – | 92.9%† |
| ARC-AGI-1 | ARC Prize | 93.5% | 94.5% |
| ARC-AGI-2 | ARC Prize | 78.1% | 60.4% |
| CritPt | Artificial Analysis | 24.0% | 23.4% |
| Chess Puzzles | Epoch AI Benchmarking Hub | 46.0% | 36.0% |
| Mystery Game Puzzles | Epoch AI Benchmarking Hub | 49.0% | 26.0% |
| LMCA | Epoch AI Benchmarking Hub | – | 52.7% |
| DTBench | Epoch AI Benchmarking Hub | – | 91.2% |
| ForecastBench | ForecastBench | – | 64.6% |
| Coding | |||
| SciCode | SciCode | – | 59.5% |
| WeirdML | WeirdML | – | 82.6% |
| FrontierCode | Cognition | – | 44.2%† |
| Arena Coding | Arena (LMArena) | – | 1539 |
| Arena WebDev | Arena (LMArena) | 1664 | 1655 |
| LiveBench 2.0 Coding | LiveBench | 96.0%† | 59.3%† |
| LiveBench 1.0 Coding | LiveBench | – | 81.5%† |
| SciCode (AA) | Artificial Analysis | 55.9% | 59.5% |
| LiveCodeBench | Vals AI | – | 87.2% |
Show all 133 benchmarks
| Metric | Publisher | DeepSeek V4.1 Flash 1017 (max) | Kimi K3 (max) |
|---|---|---|---|
| Coding | |||
| IOI | Vals AI | – | 48.9% |
| SWE-bench (Vals) | Vals AI | – | 93.4%† |
| Code Migration | Vals AI | 44.2% | 16.1% |
| Code Migration v1.1 | Vals AI | 67.5% | 28.6% |
| ProgramBench | Vals AI | – | 2.0%† |
| SWE-bench Multimodal | Vals AI | – | 44.2% |
| Vibe Code Bench 1-100 | Vals AI | – | 18.2% |
| Vibe Code Bench v1.1 | Vals AI | 87.7% | 85.0% |
| Surface Evolver Bench | Epoch AI Benchmarking Hub | – | 93.0% |
| DeepSWE v1.1 | Epoch AI Benchmarking Hub | – | 68.5% |
| ALE-Bench | Epoch AI Benchmarking Hub | – | 1524.5 |
| FrontierSWE | Epoch AI Benchmarking Hub | – | 25.9% |
| Terminal-Bench 4.0 (AA) | Artificial Analysis | 33.8% | 12.6% |
| Terminal-Bench 2.1 (AA) | Artificial Analysis | – | 85.0% |
| Startrise Frontend | Startrise | – | 79.8%† |
| FrontierCode 1.1 (AA) | Artificial Analysis | 43.7% | 43.6% |
| Agents & tools | |||
| APEX-Agents | Mercor | – | 47.8%† |
| Arena Agent | Arena (LMArena) | – | 1.8 |
| LiveBench Long Agentic | LiveBench | 50.9%† | 42.8%† |
| LiveBench Browser | LiveBench | 76.0%† | 16.0%† |
| LiveBench Agentic Coding | LiveBench | – | 62.2%† |
| Terminal-Bench 2.1 (Vals) | Vals AI | – | 80.9%† |
| MCP Atlas | Scale AI SEAL | – | 82.3% |
| HiL-Bench | Scale AI SEAL | – | 42.7%† |
| τ³-bench Banking | τ²-bench | – | 37.1% |
| GDPval-AA v2.1 | Artificial Analysis | 62.3% | – |
| τ³-Banking (AA) | Artificial Analysis | – | 46.0% |
| ITBench SRE (AA) | Artificial Analysis | – | 47.7% |
| Analyst Agent (AA) | Artificial Analysis | – | 38.8% |
| APEX-Agents (AA) | Artificial Analysis | – | 41.3% |
| CyberBench | Vals AI | – | 75.2% |
| Finance Agent v2 | Vals AI | 57.0% | 53.1% |
| Finance Agent v2.1 | Vals AI | 62.9% | 60.0% |
| Harvey's Legal Agent Benchmark v1.1 | Vals AI | 14.2% | 12.9% |
| Harvey's Legal Agent Benchmark v1 | Vals AI | 9.2% | 12.9% |
| Legal Research Bench | Vals AI | 45.2% | 46.1% |
| Tax Agent Bench | Vals AI | 44.4% | 33.4% |
| Terminal-Bench 4.0 (Vals) | Vals AI | 31.8% | 17.2% |
| Terminal-Bench Science | Vals AI | – | 1.4% |
| Time Horizon Index: KSP | Vals AI | – | 10.5% |
| PostTrainBench | Epoch AI Benchmarking Hub | – | 32.0%† |
| GBAEval | Epoch AI Benchmarking Hub | – | 48.4%† |
| Vending-Bench 2 | Andon Labs | – | 5165.0† |
| AutomationBench | Artificial Analysis | 66.6% | 58.3% |
| GDP.pdf (AA) | Artificial Analysis | 17.2% | 22.0% |
| Harvey LAB-AA v1.1 | Harvey | – | 5.3% |
| EnterpriseOps-Gym | Artificial Analysis | – | 45.3% |
| AA-Briefcase v1.1 | Artificial Analysis | 1570 | 1499 |
| Terminal-Bench-Science 0.1 (AA) | Artificial Analysis | 12.9% | 8.1% |
| CWE-Bench (AA) | Artificial Analysis | 69.2% | 53.3% |
| DeepSec-Bench (AA) | Artificial Analysis | 24.9% | 12.4% |
| CyberGym end-to-end (AA) | Artificial Analysis | 76.3% | 58.0% |
| Maths | |||
| FrontierMath Tiers 1–3 | Epoch AI Benchmarking Hub | 71.2% | 72.2% |
| FrontierMath Tier 4 | Epoch AI Benchmarking Hub | 46.3% | 39.0% |
| OTIS Mock AIME | Epoch AI Benchmarking Hub | 99.4% | 97.2% |
| ProofBench | Vals AI | – | 87.0%† |
| LiveBench Mathematics | LiveBench | – | 84.4%† |
| AIME 2026 | MathArena | – | 96.7%† |
| HMMT February 2026 | MathArena | – | 97.0%† |
| MathArena Apex | MathArena | – | 65.6%† |
| Arena Maths | Arena (LMArena) | – | 1499 |
| Knowledge | |||
| SimpleQA Verified | Epoch AI Benchmarking Hub | 42.5% | 50.6% |
| LiveBench 2.0 Data Analysis | LiveBench | 79.5%† | 78.7%† |
| LiveBench Research | LiveBench | 79.5%† | 79.0%† |
| LiveBench 1.0 Data Analysis | LiveBench | – | 78.7%† |
| AA-Omniscience | Artificial Analysis | 4.7 | 19.7 |
| MMLU-Pro | Vals AI | – | 88.0% |
| LegalBench | Vals AI | – | 86.2% |
| CorpFin | Vals AI | – | 71.6%† |
| TaxEval | Vals AI | – | 75.7%† |
| BioMysteryBench | Vals AI | – | 72.6% |
| Excel Modeling Benchmark | Vals AI | 61.8% | 66.7% |
| Excel Modeling Benchmark v1.1 | Vals AI | 69.2% | 71.3% |
| MedCode | Vals AI | – | 49.4% |
| MedScribe | Vals AI | – | 88.0% |
| Public Benefits Bench v1.1 | Vals AI | – | 68.2% |
| GDP.pdf | Epoch AI Benchmarking Hub | – | 19.0% |
| AA-Omniscience: accuracy | Artificial Analysis | 42.9% | 47.6% |
| AA-Omniscience: non-hallucination | Artificial Analysis | 33.3% | 46.8% |
| Instruction following | |||
| LiveBench Language | LiveBench | – | 85.5%† |
| LiveBench Instruction Following | LiveBench | – | 71.4%† |
| Arena Instruction Following | Arena (LMArena) | – | 1489 |
| Human preference | |||
| Arena Text | Arena (LMArena) | – | 1487 |
| EQ-Bench 4 | EQ-Bench | – | 1339† |
| Arena Creative Writing | Arena (LMArena) | – | 1462 |
| Arena Multi-turn | Arena (LMArena) | – | 1502 |
| Arena Longer Queries | Arena (LMArena) | – | 1503 |
| Multimodal | |||
| LiveBench Design | LiveBench | 83.4%† | 77.2%† |
| MMMU-Pro | Artificial Analysis | 80.6% | 80.5% |
| MMMU-Pro (Vals) | Vals AI | – | 88.2%† |
| MortgageTax | Vals AI | – | 66.3%† |
| SAGE | Vals AI | – | 52.8% |
| Blueprint-Bench 2 | Epoch AI Benchmarking Hub | – | 29.5%† |
| AA-3D | Artificial Analysis | – | 981 |
| Long context | |||
| AA-LCR | Artificial Analysis | 82.7% | 88.7% |
| MLCR | Artificial Analysis | 26.7% | 21.7% |
| Composite | |||
| Epoch Capabilities Index | Epoch AI Benchmarking Hub | – | 157.9† |
| LiveBench 2.0 | LiveBench | 79.2%† | 63.4%† |
| LiveBench 1.0 | LiveBench | – | 79.2%† |
| AA Intelligence Index v4.3.2 | Artificial Analysis | 45.1 | 43.6 |
| Vals Index | Vals AI | 61.3 | 53.0 |
| Vals Finance Index | Vals AI | 60.9% | 58.0% |
| Vals RSI Index | Vals AI | – | 20.9% |
| Vals Multimodal Index | Vals AI | – | 73.4%† |
| Vals Router Index | Vals AI | 61.3%† | 53.0%† |
| AA Cyber Index | Artificial Analysis | 56.8 | 41.3 |
| Safety & honesty | |||
| FORTRESS | Scale AI SEAL | – | 26.6% |
| Metadata | |||
| AA Openness Index | Artificial Analysis | 44.4 | 38.9 |
Data as of 2026-10-24. Best configuration of each model; every score links to its source on the model pages.
What changed
- DeepSeek V4.1 Flash 1017 (max): 10 new results (MLCR, AA Cyber Index, CWE-Bench (AA)…).
- Kimi K3 (max): 7 new results (Code Migration v1.1, Excel Modeling Benchmark v1.1, Finance Agent v2.1…).
- DeepSeek V4.1 Flash 1017 (max): BenchLeader Index 63.1 → 63.9.
- DeepSeek V4.1 Flash 1017 (max): 11 new results (GPQA Diamond, FrontierMath Tiers 1–3, FrontierMath Tier 4…).
- Kimi K3 (max): 1 new result (Terminal-Bench-Science 0.1 (AA)); 1 score revision.
- DeepSeek V4.1 Flash 1017 (max): 1 new result (FrontierCode 1.1 (AA)); 2 score revisions.
- Kimi K3 (max): 3 score revisions.
- Kimi K3 (max): 1 score revision.
- Kimi K3 (max): 1 new result (τ³-bench Banking).
- Kimi K3 (max): 8 new results (LegalBench, Vals Index, Finance Agent v2…); 5 score revisions.
Related comparisons
- DeepSeek V4.1 Flash 1017 vs Claude Opus 5.5
- DeepSeek V4.1 Flash 1017 vs Gemini 4 Argon
- DeepSeek V4.1 Flash 1017 vs Muse Wonder Full
- Kimi K3 vs Claude Opus 5.5
- Kimi K3 vs Gemini 4 Argon
- Kimi K3 vs Muse Wonder Full
- Kimi K3 vs Claude Fable 5.1
- DeepSeek V4.1 Flash 1017 vs GPT-6 Astra
Rankings: Best LLM overall · Best LLM for agents · Best cheap LLMs
DeepSeek V4.1 Flash 1017 vs Kimi K3: questions
- Is DeepSeek V4.1 Flash 1017 better than Kimi K3?
- DeepSeek V4.1 Flash 1017 (max) and Kimi K3 (max) are neck and neck on quality (64.0 vs 64.7, within the ±5.1 margin). The BenchLeader Index combines every independent quality benchmark; the two are within the margin of error as of 2026-10-24, so check the category scores for your use.
- Which is cheaper, DeepSeek V4.1 Flash 1017 or Kimi K3, and will that change?
- DeepSeek V4.1 Flash 1017 is cheaper today: $0.525 against $6.00 per million tokens, blended at three input tokens per output token. No price change has been announced for either.
- Is DeepSeek V4.1 Flash 1017 better than Kimi K3 for coding?
- DeepSeek V4.1 Flash 1017 scores higher in coding (70 vs 65 on the category index, where 50 is average). On Arena WebDev, the top shared coding benchmark, it is 1664 for DeepSeek V4.1 Flash 1017 against 1655 for Kimi K3.
- Is DeepSeek V4.1 Flash 1017 better than Kimi K3 for agents and tool use?
- DeepSeek V4.1 Flash 1017 scores higher in agentic tasks (63 vs 59 on the category index, where 50 is average). On CWE-Bench (AA), the top shared agents & tools benchmark, it is 69.2% for DeepSeek V4.1 Flash 1017 against 53.3% for Kimi K3.
- Which is better for long documents?
- Kimi K3 scores higher on long-context benchmarks (70 vs 66). Both accept 1,048,576 tokens of context.
- Which has lower latency?
- DeepSeek V4.1 Flash 1017 (max) answers first: 11.2 s against 55.4 s to the first answer token, as Artificial Analysis measures it. Reasoning effort drives most of this: a higher setting thinks longer before it answers.
- Which is faster, DeepSeek V4.1 Flash 1017 or Kimi K3?
- DeepSeek V4.1 Flash 1017 streams faster: 198 against 39 output tokens per second.
- Do independent benchmarks agree on DeepSeek V4.1 Flash 1017 vs Kimi K3?
- Across the 43 benchmarks both have, DeepSeek V4.1 Flash 1017 leads 29, Kimi K3 leads 14. Every overall index that rates both puts DeepSeek V4.1 Flash 1017 ahead.
- When was DeepSeek V4.1 Flash 1017 released?
- DeepSeek V4.1 Flash 1017 was released on 17 Oct 2026.