DeepSeek-V4.1-FlashvsKimi K3
在 28 个同模式 benchmark 中，DeepSeek-V4.1-Flash 整体领先：DeepSeek-V4.1-Flash 领先 15 项，Kimi K3 领先 12 项，持平 1 项，平均分差 -19.36。另有 20 项测试模式不同，仅供参考。
DeepSeek-V4.1-Flash
DeepSeek-AI · 2026-09-10 · 多模态大模型
Kimi K3
Moonshot AI · 2026-07-16 · 推理大模型
DeepSeek-V4.1-Flash15 项(54%)持平1(43%)12 项Kimi K3
评测分数
按能力类目分组，每组内按分差大小排列；共 28 项。
数学
Kimi K3 领先 3/4| 评测项 | DeepSeek-V4.1-Flash | Kimi K3 | 分差 |
|---|---|---|---|
| Riemann-bench | 2425 / 42Max (With Tools + Internet) | 37.6017 / 42Max (With Tools + Internet) | -13.60 |
| FrontierMath Tier 4 v2 | 26.8344 / 77Max (With Tools) | 39.0233 / 77Max (With Tools) | -12.20 |
| FrontierMath v2 | 67.3732 / 119Max (With Tools) | 72.1825 / 119Max (With Tools) | -4.81 |
| MathArena Apex | 65.604 / 17Max (No Tools) | 65.604 / 17Max (No Tools) | 持平 |
仓库修复与多文件工程
DeepSeek-V4.1-Flash 领先 3/3| 评测项 | DeepSeek-V4.1-Flash | Kimi K3 | 分差 |
|---|---|---|---|
| NL2Repo-Bench | 65.403 / 20Max (With Tools) | 588 / 20Max (With Tools) | +7.40 |
| DeepSWE | 74.208 / 108Max (With Tools) | 68.5136 / 108Max (With Tools) | +5.69 |
| APEX-SWE | 51.9014 / 72Max (With Tools) | 4818 / 72Max (With Tools) | +3.90 |
图像感知与视觉推理
DeepSeek-V4.1-Flash 领先 2/3| 评测项 | DeepSeek-V4.1-Flash | Kimi K3 | 分差 |
|---|---|---|---|
| ZeroBench Main | 4910 / 17Max (With Tools) | 4112 / 17Max (With Tools) | +8 |
| BabyVision | 89.609 / 17Max (With Tools) | 85.7011 / 17Max (With Tools) | +3.90 |
| MMMU-Pro | 7779 / 237Max (No Tools) | 80.5045 / 237Max (No Tools) | -3.50 |
自主开发与终端任务
DeepSeek-V4.1-Flash 领先 2/3| 评测项 | DeepSeek-V4.1-Flash | Kimi K3 | 分差 |
|---|---|---|---|
| Terminal-Bench 4.0 | 26.8044 / 110Max (With Tools) | 12.6064 / 110Max (With Tools) | +14.20 |
| Terminal-Bench 3.0 | 304 / 11Max (With Tools) | 17.708 / 11Max (With Tools) | +12.30 |
| Vibe Code Bench 1-100 | 16.3816 / 25Max (With Tools) | 18.2414 / 25Max (With Tools) | -1.86 |
文档与图表理解
胶着 2/2| 评测项 | DeepSeek-V4.1-Flash | Kimi K3 | 分差 |
|---|---|---|---|
| Chartography | 78.9010 / 19Max (With Tools) | 68.1013 / 19Max (With Tools) | +10.80 |
| GDP.pdf | 12.8074 / 126Max (No Tools) | 2234 / 126Max (No Tools) | -9.20 |
科学计算与科研编程
胶着 2/2| 评测项 | DeepSeek-V4.1-Flash | Kimi K3 | 分差 |
|---|---|---|---|
| Terminal-Bench-Science 0.1 | 15.709 / 19Max (With Tools) | 7.1015 / 19Max (With Tools) | +8.60 |
| SciCode | 51.9063 / 136Max (No Tools) | 59.508 / 136Max (No Tools) | -7.60 |
跨能力综合测试
DeepSeek-V4.1-Flash 领先 2/2| 评测项 | DeepSeek-V4.1-Flash | Kimi K3 | 分差 |
|---|---|---|---|
| LiveBench 2.0 | 78.0412 / 49Max (task-specific tools) | 63.3934 / 49Max (task-specific tools) | +14.65 |
| SuperCLUE | 71.812 / 13Reported best (effort unspecified) | 70.685 / 13Reported best (effort unspecified) | +1.13 |
写作与创意表达
Kimi K3 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | Kimi K3 | 分差 |
|---|---|---|---|
| Creative Writing | 1,54053 / 118Normal (No Tools) | 2,0845 / 118Normal (No Tools) | -543.60 |
办公与文档流程
DeepSeek-V4.1-Flash 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | Kimi K3 | 分差 |
|---|---|---|---|
| AutomationBench | 54.806 / 148Max (With Tools) | 46.7018 / 148Max (With Tools) | +8.10 |
机器学习工程
Kimi K3 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | Kimi K3 | 分差 |
|---|---|---|---|
| WeirdML v3 | 0.0620 / 24Thinking High (With Tools) | 0.0816 / 24Thinking High (With Tools) | -0.02 |
漏洞发现与分析
DeepSeek-V4.1-Flash 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | Kimi K3 | 分差 |
|---|---|---|---|
| CyberGym | 88.103 / 14Max (With Tools) | 809 / 14Max (With Tools) | +8.10 |
科学知识与推理
Kimi K3 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | Kimi K3 | 分差 |
|---|---|---|---|
| CritPt | 14.3063 / 207Max (No Tools) | 23.4031 / 207Max (No Tools) | -9.10 |
程序生成与编辑
DeepSeek-V4.1-Flash 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | Kimi K3 | 分差 |
|---|---|---|---|
| Program Bench | 20.3012 / 18Max (With Tools) | 17.5015 / 18Max (With Tools) | +2.80 |
综合知识考试
DeepSeek-V4.1-Flash 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | Kimi K3 | 分差 |
|---|---|---|---|
| HLE | 63.906 / 247Max (With Tools) | 59.8013 / 247Max (With Tools) | +4.10 |
跨应用与工具编排
Kimi K3 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | Kimi K3 | 分差 |
|---|---|---|---|
| APEX-Agents | 39.5049 / 63Reported best (effort unspecified) | 47.8034 / 63Reported best (effort unspecified) | -8.30 |
金融
Kimi K3 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | Kimi K3 | 分差 |
|---|---|---|---|
| Tax Agent Bench | 26.6345 / 72Max (With Tools) | 68.677 / 72Max (With Tools) | -42.04 |
测试模式不同的成绩
共 20 项。两款模型的公开成绩来自不同测试模式，分数不直接可比，不计入胜负和平均分差。
| 评测项 | DeepSeek-V4.1-Flash | Kimi K3 |
|---|---|---|
| ARC-AGI-1 | 7.83Normal (No Tools) | 94.50Max (With Tools) |
| IOI (Vals v2) | 40.28Thinking High (With Tools) | 48.94Max (With Tools) |
| BioMysteryBench | 67.78Thinking High (With Tools) | 72.59Max (With Tools) |
| Vals Index | 57.86Thinking High (With Tools) | 57.81Max (With Tools) |
| Vals Index v2.1 | 51.32Thinking High (With Tools) | 50.70Reported best (effort unspecified) |
| MedCode | 41.17Thinking High (With Tools) | 48.88Max (With Tools) |
| MedScribe | 85.50Thinking High (With Tools) | 87.96Max (With Tools) |
| FrontierCode 1.1 Main | 43.90Thinking High (With Tools) | 44.20Normal (With Tools) |
| Vibe Code Bench v1.1 | 84.74Thinking High (With Tools) | 84.96Max (With Tools) |
| SimpleBench | 66.70Historical report (mode unspecified) | 60.70Max (No Tools) |
| LiveBench | 81.11Max (No Tools) | 79.19Reported best (effort unspecified) |
| EMB (Excel Modeling) | 57.21Thinking High (With Tools) | 66.40Max (With Tools) |
| Finance Agent v2 | 53.48Thinking High (With Tools) | 54.40Max (With Tools) |
| Harvey's Legal Agent Benchmark | 6.67Thinking High (With Tools) | 12.92Max (With Tools) |
| Legal Research Bench | 41.35Thinking High (With Tools) | 44.20Max (With Tools) |
| Code Migration | 45.62Thinking High (With Tools) | 16.10Max (With Tools) |
| ProofBench v1.1 (Lean 4) | 54Thinking High (With Tools) | 87Max (With Tools) |
| GPQA Diamond | 90.90Max (No Tools) | 91.92Thinking High (No Tools) |
| Public Benefits Bench v1.1 | 64.28Thinking High (With Tools) | 68.27Max (With Tools) |
| SAGE | 47.88Thinking High (With Tools) | 54.26Max (With Tools) |
规格对比
| 字段 | DeepSeek-V4.1-Flash | Kimi K3 |
|---|---|---|
| 发布机构 | DeepSeek-AI | Moonshot AI |
| 发布时间 | 2026-09-10 | 2026-07-16 |
| 模型类型 | 多模态大模型 | 推理大模型 |
| 架构 | MoE 架构 | MoE 架构 |
| 参数规模 | 5520亿 | 2.8万亿 |
| 上下文长度 | 1M | 1M |
| 最大输出 | 384K | 1M |
API 调用价格
价格优先使用 DataLearner 配置的 API 记录；缺失项不做推测。
| 价格项 | DeepSeek-V4.1-Flash | Kimi K3 |
|---|---|---|
| 文本输入 | ¥1 / 1M tokens | ¥20 / 1M tokens |
| 文本输出 | ¥4 / 1M tokens | ¥100 / 1M tokens |
| 缓存读取 | ¥0.02 / 1M tokens | ¥2 / 1M tokens |
| 缓存写入 | 暂无公开价格 | ¥20 / 1M tokens |
小结
- DeepSeek-V4.1-Flash在以下类目领先:仓库修复与多文件工程 (3/3)、图像感知与视觉推理 (2/3)、自主开发与终端任务 (2/3)、跨能力综合测试 (2/2)、办公与文档流程 (1/1)、漏洞发现与分析 (1/1)、程序生成与编辑 (1/1)、综合知识考试 (1/1)
- Kimi K3在以下类目领先:数学 (3/4)、写作与创意表达 (1/1)、机器学习工程 (1/1)、科学知识与推理 (1/1)、跨应用与工具编排 (1/1)、金融 (1/1)
- 胶着类目:文档与图表理解、科学计算与科研编程
28 个同模式 benchmark 上，Kimi K3 平均高出 19.36 分。
单项差距最大的 benchmark：Creative Writing — DeepSeek-V4.1-Flash 1,540，Kimi K3 2,084（分差 -543.60）。
本页正文由结构化模型、价格与 benchmark 数据生成，不使用实时 LLM 撰写。