Introduction
In late July 2026, two flagship Chinese models released benchmark data almost simultaneously:
- DeepSeek-V4-Flash-0731 — DeepSeek's efficiency inference variant (speculative decoding enhanced)
- Kimi-K3 — Moonshot's MoE flagship (2.8T total params / 114B activated)
Data sources: DeepSeek-V4-Flash-0731 Model Card · Kimi-K3 Model Card (collected 2026-08-01)
Part 1: In-Depth Comparison
1.1 Model Positioning & Architecture Differences
This is the essential context — these are not same-class competitors:
| Dimension | DeepSeek-V4-Flash-0731 | Kimi-K3 |
|---|---|---|
| Positioning | Efficient inference Flash variant | Flagship all-rounder |
| Architecture | Speculative decoding enhanced (DSpark) | MoE (1 Dense + 69 KDA + 24 Gated MLA) |
| Total params | Undisclosed (Flash variant, emphasizes small activation) | 2.8T |
| Activated params | Far smaller than Pro variant | 104B |
| Experts | — | 896 (top-16 + 2 shared) |
| Context | — | 1M tokens |
| Multimodal | Not reflected in benchmarks | Vision benchmarks included |
| Baselines | V4-Pro Preview, GLM-5.2, Opus-4.8 | Claude Fable 5, GPT-5.6 Sol, Opus-4.8, GPT-5.5 |
Core difference: DeepSeek-V4-Flash is the "small and refined" efficiency variant pursuing cost-performance; Kimi-K3 is the "large and comprehensive" flagship covering reasoning/coding/agent/vision.
1.2 Shared Benchmark Head-to-Head
Both models have results on 5 shared benchmarks. Kimi-K3 leads on all 5.
| Benchmark | DeepSeek-V4-Flash-0731 | Kimi-K3 (max) | Delta | Winner |
|---|---|---|---|---|
| Terminal Bench 2.1 | 82.7 | 88.3 | +5.6 | Kimi-K3 |
| DeepSWE | 54.4 | 67.5 | +13.1 🏆 | Kimi-K3 |
| Toolathlon-Verified | 70.3 | 76.5 | +6.2 | Kimi-K3 |
| Agents' Last Exam | 25.2 | 28.3 | +3.1 | Kimi-K3 |
| AutomationBench | 25.1 (Public) | 30.8 | +5.7 | Kimi-K3 |
🏆 = largest gap
Gap breakdown:
| Level | Benchmark | Delta | Interpretation |
|---|---|---|---|
| 🔴 Large lead | DeepSWE | +13.1 | Kimi dominates SWE code reasoning |
| 🟡 Medium lead | Terminal Bench | +5.6 | Kimi stronger at terminal operations |
| 🟡 Medium lead | Toolathlon | +6.2 | Kimi better at tool use |
| 🟡 Medium lead | AutomationBench | +5.7 | Kimi better at automation |
| 🟢 Small lead | Agents' Last Exam | +3.1 | Smallest gap, still Kimi |
1.3 Indirect Comparison via GLM-5.2 Anchor
Both models benchmarked GLM-5.2 as an indirect reference. Note that GLM-5.2 scores differ between cards (evaluation configs may vary).
| Benchmark | DeepSeek-V4-Flash | Kimi-K3 | Winner | GLM-5.2 (DeepSeek card) | GLM-5.2 (Kimi card) |
|---|---|---|---|---|---|
| Terminal Bench 2.1 | 82.7 | 88.3 | Kimi +5.6 | 81.0 | 82.7 |
| DeepSWE | 54.4 | 67.5 | Kimi +13.1 | 46.2 | 46.2 |
| Toolathlon-Verified | 70.3 | 76.5 | Kimi +6.2 | 59.9 | 59.9 |
| Agents' Last Exam | 25.2 | 28.3 | Kimi +3.1 | 23.8 | 20.4 |
| AutomationBench | 25.1 (Public) | 30.8 | Kimi +5.7 | 12.9 | 12.9 |
Key findings:
- In DeepSeek's card, Flash-0731 (82.7) roughly ties GLM-5.2 (81.0) on Terminal Bench, while Kimi (88.3) far exceeds both
- On DeepSWE, both crush GLM-5.2 (46.2), but Kimi (67.5) beats DeepSeek (54.4) by 13.1 points
- Agents' Last Exam has inconsistent GLM-5.2 scores across cards (23.8 vs 20.4)
1.4 DeepSeek-V4-Flash-0731 Exclusive Benchmarks
DeepSeek evaluated additional Code Agent benchmarks:
| Benchmark | Flash-0731 | Flash (Preview) | V4-Pro (Preview) | GLM-5.2 | Opus-4.8 | Flash vs Opus |
|---|---|---|---|---|---|---|
| NL2Repo | 54.2 | 39.4 | 38.5 | 48.9 | 69.7 | −15.5 |
| Cybergym | 76.7 | 38.7 | 52.7 | — | 83.1 | −6.4 |
| DSBench-FullStack † | 68.7 | 37.0 | 41.8 | 61.8 | 71.6 | −2.9 |
| DSBench-Hard † | 59.6 | 25.8 | 31.1 | 54.5 | 71.7 | −12.1 |
Flash-0731 leads its own Preview on all 4 exclusive benchmarks, but trails Opus-4.8 on all of them
1.5 Kimi-K3 Exclusive Benchmark Highlights
Kimi-K3 covers far more benchmarks across reasoning, coding, agent, and vision.
Reasoning & Knowledge
Kimi strengths: AA-LCR · Kimi weaknesses: CritPt (gap 8.9)
| Benchmark | Kimi-K3 | Top Competitor | Competitor Score | Result |
|---|---|---|---|---|
| GPQA Diamond | 93.5 | GPT-5.6 Sol | 94.1 | 🥈 Lost by 0.6 |
| CritPt | 23.4 | GPT-5.6 Sol | 32.3 | 🥉 Gap 8.9 |
| AA-LCR | 74.7 | Kimi wins | — | 🏆 Champion |
| HLE-Full | 43.5 / 56.0 | Claude Fable 5 | 53.3 / 63.0 | 🥈 Lost by 9.8 |
Coding
Kimi strengths: ProgramBench / SWE-Marathon · Kimi weaknesses: DeepSWE / FrontierSWE
| Benchmark | Kimi-K3 | Top Competitor | Competitor Score | Result |
|---|---|---|---|---|
| DeepSWE | 67.5 | GPT-5.6 Sol | 73.0 | 🥈 Gap 5.5 |
| ProgramBench | 77.8 | Kimi wins | — | 🏆 Champion |
| Terminal-Bench 2.1 | 88.3 | GPT-5.6 Sol | 88.8 | 🥈 Lost by 0.5 |
| FrontierSWE | 81.2 | Claude Fable 5 | 86.6 | 🥈 Gap 5.4 |
| SWE-Marathon | 42.0 | Kimi wins | — | 🏆 Champion (GPT-5.5 only 14.0) |
Agent
Kimi strengths: BrowseComp / MCPMark / DeepSearchQA / Harvey · Kimi weaknesses: GDPval
| Benchmark | Kimi-K3 | Top Competitor | Competitor Score | Result |
|---|---|---|---|---|
| BrowseComp | 91.2 | Kimi wins | — | 🏆 Champion |
| MCPMark-Verified | 94.5 | Kimi wins | — | 🏆 Champion |
| DeepSearchQA F1 | 95.0 | Kimi wins | — | 🏆 Champion |
| Harvey Lab-AA | 94.6 | Kimi wins | — | 🏆 Champion |
| GDPval-AA v2 (Elo) | 1686 | Claude Fable 5 | 1747 | 🥈 Lost by 61 |
Vision
Kimi strengths: OmniDocBench / Video-MME / MMVU · Kimi weaknesses: WorldVQA / BabyVision
| Benchmark | Kimi-K3 | Top Competitor | Competitor Score | Result |
|---|---|---|---|---|
| OmniDocBench | 91.1 | Kimi wins | — | 🏆 Champion |
| Video-MME (w. sub) | 90.0 | Kimi wins | — | 🏆 Champion |
| MathVision | 94.3 / 97.8 | GPT-5.6 Sol | 95.8 / 97.8 | 🥈 Lost by 1.5 |
| MMVU | 82.1 | Kimi wins | — | 🏆 Champion |
| WorldVQA ForceAnswer | 51.0 | Claude Fable 5 | 56.7 | 🥈 Gap 5.7 |
| BabyVision w/ python | 85.7 | Claude Fable 5 | 90.5 | 🥈 Gap 4.8 |
1.6 Evaluation Methodology Differences (Important)
⚠️ Evaluation configs may differ between models — direct comparison requires caution.
| Dimension | DeepSeek-V4-Flash-0731 | Kimi-K3 |
|---|---|---|
| Reasoning effort | max (temperature=1.0, top_p=0.95) | max |
| Agent framework | DeepSeek Harness minimal mode | Undisclosed |
| AutomationBench version | Public subset | Full version |
| Benchmark count | 9 | 40+ |
1.7 Summary
| Dimension | Conclusion | Evidence |
|---|---|---|
| Shared benchmarks | Kimi-K3 leads 5/5 | DeepSWE +13.1 · Terminal +5.6 · Toolathlon +6.2 |
| Model scale | Asymmetric comparison | Kimi 2.8T/104B vs Flash undisclosed (small activation) |
| Cost-performance | Flash punches above its weight | Flash Terminal Bench (82.7) ≈ GLM-5.2 (81.0) |
| Versatility | Kimi covers reasoning+coding+agent+vision | 40+ benchmarks, 15+ 🏆 champions |
| Coding | Kimi leads by largest margin | DeepSWE 67.5 vs 54.4 (+13.1) |
| Agent | Kimi stronger at tool use | Toolathlon +6.2 · AutomationBench +5.7 |
One-liner: Kimi-K3 is today's most versatile multimodal flagship; DeepSeek-V4-Flash-0731 proves that DSpark speculative decoding can approach flagship-level Code Agent capability with minimal activation.
Part 2: Raw Benchmark Data
Below are the unprocessed benchmark figures from each model's HuggingFace card.
2.1 DeepSeek-V4-Flash-0731
Positioning: Efficient inference Flash variant with speculative decoding module (DeepSeek-V4-Flash-DSpark)
Source: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731
| Benchmark | DeepSeek-V4-Flash-0731 | DeepSeek-V4-Flash (Preview) | DeepSeek-V4-Pro (Preview) | GLM-5.2 | Opus-4.8 |
|---|---|---|---|---|---|
| Terminal Bench 2.1 | 82.7 | 61.8 | 72.1 | 81.0 | 85.0 |
| NL2Repo | 54.2 | 39.4 | 38.5 | 48.9 | 69.7 |
| Cybergym | 76.7 | 38.7 | 52.7 | - | 83.1 |
| DeepSWE | 54.4 | 7.3 | 12.8 | 46.2 | 58.0 |
| Toolathlon-Verified | 70.3 | 49.7 | 55.9 | 59.9 | 76.2 |
| Agents' Last Exam | 25.2 | 15.8 | 16.5 | 23.8 | 25.7 |
| AutomationBench Public | 25.1 | 10.8 | 12.8 | 12.9 | 27.2 |
| DSBench-FullStack † | 68.7 | 37.0 | 41.8 | 61.8 | 71.6 |
| DSBench-Hard † | 59.6 | 25.8 | 31.1 | 54.5 | 71.7 |
† DeepSeek internal benchmark
2.2 Kimi-K3
Positioning: MoE flagship model
Architecture Specs:
| Spec | Value |
|---|---|
| Total params | 2.8T |
| Activated params | 104B |
| Layers | 93 (1 Dense + 69 KDA + 24 Gated MLA) |
| Experts | 896 (top-16 + 2 shared) |
| Context length | 1M tokens |
Source: https://huggingface.co/moonshotai/Kimi-K3
Reasoning & Knowledge
| Benchmark | Kimi K3 (max) | Claude Fable 5 (max) | GPT-5.6 Sol (max) | Claude Opus 4.8 (max) | GPT-5.5 (xhigh) | GLM-5.2 (max) |
|---|---|---|---|---|---|---|
| GPQA Diamond | 93.5 | 92.6 | 94.1 | 91.0 | 93.5 | 91.2 |
| CritPt | 23.4 | 28.6 | 32.3 | 20.9 | 27.1 | 20.9 |
| AA-LCR | 74.7 | 70.0 | 73.7 | 67.7 | 74.3 | 71.3 |
| HLE-Full | 43.5 / 56.0 | 53.3 / 63.0 | 44.5 / 58.0 | 49.8 / 57.9 | 41.4 / 52.2 | — |
Coding
| Benchmark | Kimi K3 (max) | Claude Fable 5 (max) | GPT-5.6 Sol (max) | Claude Opus 4.8 (max) | GPT-5.5 (xhigh) | GLM-5.2 (max) |
|---|---|---|---|---|---|---|
| DeepSWE | 67.5 | 70.0 | 73.0 | 59.0 | 67.0 | 46.2 |
| ProgramBench | 77.8 | 76.8 | 77.6 | 71.9 | 70.8 | 63.7 |
| Terminal-Bench 2.1 | 88.3 | 88.0 | 88.8 | 84.6 | 83.4 | 82.7 |
| FrontierSWE | 81.2 | 86.6 | 71.3 | 66.7 | 64.9 | 67.3 |
| SWE-Marathon | 42.0 | 35.0 | 39.0 | 40.0 | 14.0 | 13.0 |
| PostTrainBench | 36.6 | 41.4 | 34.6 | 34.1 | 28.4 | 34.3 |
| MLS-Bench-Lite | 48.3 | 49.9 | 46.2 | 42.8 | 35.5 | 40.4 |
| SciCode | 58.7 | 60.2 | 56.1 | 53.5 | 56.1 | 50.5 |
| Kimi Code Bench 2.0 | 72.9 | 76.9 | 64.8 | 71.7 | 69.0 | 64.2 |
Agentic
| Benchmark | Kimi K3 (max) | Claude Fable 5 (max) | GPT-5.6 Sol (max) | Claude Opus 4.8 (max) | GPT-5.5 (xhigh) | GLM-5.2 (max) |
|---|---|---|---|---|---|---|
| BrowseComp | 91.2 | 88.0 | 90.4 | 84.3 | 84.4 | — |
| DeepSearchQA (F1) | 95.0 | 94.2 | — | 93.1 | — | — |
| ResearchRubrics | 76.2 | — | 73.8 | 73.5 | 64.0 | 71.1 |
| GDPval-AA v2 (Elo) | 1686 | 1747 | 1736 | 1593 | 1491 | 1510 |
| Toolathlon-Verified | 76.5 | 77.9 | 74.9 | 76.2 | 73.5 | 59.9 |
| MCPMark-Verified | 94.5 | 87.4 | 92.9 | 76.4 | 92.9 | — |
| MCP-Atlas | 84.2 | 84.7 | 83.6 | 83.6 | 82.8 | 82.6 |
| AutomationBench | 30.8 | 29.1 | 29.7 | 27.2 | 22.7 | 12.9 |
| JobBench | 54.3 | 57.4 | 45.4 | 48.4 | 38.3 | 43.4 |
| AA-Briefcase (Elo) | 1548 | 1583 | 1495 | 1354 | 1158 | 1260 |
| Agents' Last Exam | 28.3 | 25.7† | 29.6 | 27.0 | 26.6 | 20.4 |
| APEX-Agents | 41.0 | 43.3 | 39.9 | 39.4 | 38.5 | 35.6 |
| OfficeQA Pro | 63.3 | 69.9 | 63.2 | 63.9 | 60.9 | 41.4 |
| SpreadsheetBench 2 | 34.8 | 34.7 | 32.4 | 31.6 | 29.1 | 28.1 |
| OSWorld-Verified | 84.8 | 85.0 | 83.0 | 83.4 | 79.0 | — |
| OSWorld 2.0 | 58.3 | 66.1 | 62.6 | 55.7 | 49.5 | — |
| SaaS-Bench | 60.1 | — | 61.4 | 56.1 | 43.8 | — |
| τ³-Banking | 33.4 | 26.8 | 33.0 | 27.6 | 31.3 | 26.8 |
| Harvey Lab-AA | 94.6 | 93.6 | 87.2 | 91.1 | 86.3 | 91.0 |
| CorpFin v2 | 71.6 | 71.8 | 64.4 | 66.7 | 68.4 | 66.1 |
| Finance Agent v2 | 54.4 | 56.3 | 53.8 | 53.9 | 51.8 | 49.7 |
| Legal Research Bench | 44.2 | 49.5 | 48.1 | 43.8 | 40.4 | 31.3 |
Vision
| Benchmark | Kimi K3 (max) | Claude Fable 5 (max) | GPT-5.6 Sol (max) | Claude Opus 4.8 (max) | GPT-5.5 (xhigh) | GLM-5.2 (max) |
|---|---|---|---|---|---|---|
| WorldVQA ForceAnswer | 51.0 | 56.7 | 41.8 | 39.1 | 38.5 | — |
| OmniDocBench | 91.1 | 89.8 | 85.8 | 87.9 | 89.4 | — |
| PerceptionBench | 58.5 | 57.2 | 59.7 | 47.2 | 55.8 | — |
| Video-MME (w. sub) | 90.0 | — | 89.5 | 86.0 | 89.3 | — |
| MMVU | 82.1 | — | 81.2 | 79.2 | 81.7 | — |
| BabyVision w/ python | 85.7 | 90.5 | 88.9 | 81.2 | 83.6 | — |
| MMMU-Pro | 81.6 / 83.4 | 81.2 / 86.5 | 83.0 / 84.6 | 78.9 / 82.7 | 81.2 / 83.2 | — |
| CharXiv (RQ) | 84.8 / 91.3 | 88.9 / 93.5 | 84.6 / 89.1 | 80.5 / 89.9 | 84.1 / 89.0 | — |
| MathVision | 94.3 / 97.8 | 94.8 / 98.6 | 95.8 / 97.8 | 86.7 / 97.1 | 92.2 / 96.8 | — |
| ZeroBench (pass@5) | 23.0 / 41.0 | 23.0 / 46.0 | 17.0 / 35.0 | 17.0 / 34.0 | 22.0 / 41.0 | — |
Data Sources
| Source | URL | Collected |
|---|---|---|
| DeepSeek-V4-Flash-0731 Model Card | https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731 | 2026-08-01 |
| Kimi-K3 Model Card | https://huggingface.co/moonshotai/Kimi-K3 | 2026-08-01 |
⚠️ Disclaimer: All data is self-reported from the models' HuggingFace cards. Evaluation configs may differ (agent framework, temperature, benchmark subsets). Interpret comparisons with caution. Click the links above for full model card details.