Hy4 previewvsKimi K3
在 16 个共同 benchmark 中，Kimi K3 整体领先：Hy4 preview 领先 3 项，Kimi K3 领先 13 项，持平 0 项，平均分差 -5.91。
Hy4 preview
腾讯AI实验室 · 2026-08-28 · 推理大模型
Kimi K3
Moonshot AI · 2026-07-16 · 推理大模型
Hy4 preview3 项(19%)(81%)13 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 16 项。
AI Agent - Tool Usage
Kimi K3 领先 3/4| 评测项 | Hy4 preview | Kimi K3 | 分差 |
|---|---|---|---|
| Terminal-Bench 2.1 | 85.4010 / 47Thinking High (With Tools) | 88.302 / 47Max (With Tools) | -2.90 |
| Toolathlon-Verified | 74.103 / 8Thinking High (With Tools) | 76.502 / 8Max (With Tools) | -2.40 |
| AutomationBench | 32.103 / 10Thinking High (With Tools) | 30.805 / 10Max (With Tools) | +1.30 |
| MCP-Atlas | 83.705 / 40Thinking High (With Tools) | 84.204 / 40Max (With Tools) | -0.50 |
Coding and Software Engineer
Kimi K3 领先 4/4| 评测项 | Hy4 preview | Kimi K3 | 分差 |
|---|---|---|---|
| Program Bench | 17.506 / 6Thinking High (With Tools) | 77.801 / 6Max (With Tools) | -60.30 |
| SWE-Marathon | 31.903 / 5Thinking High (With Tools) | 422 / 5Max (With Tools) | -10.10 |
| DeepSWE | 64.3011 / 31Thinking High (With Tools) | 67.505 / 31Max (With Tools) | -3.20 |
| PostTrain Bench | 35.604 / 5Thinking High (With Tools) | 36.603 / 5Max (With Tools) | -1 |
Agent Level Benchmark
Kimi K3 领先 2/3| 评测项 | Hy4 preview | Kimi K3 | 分差 |
|---|---|---|---|
| Job Bench | 61.701 / 3Thinking High (With Tools) | 54.303 / 3Max (With Tools) | +7.40 |
| Agents' Last Exam | 22.8013 / 14Thinking High (With Tools) | 28.305 / 14Max (With Tools) | -5.50 |
| APEX-Agents | 37.106 / 6Thinking High (With Tools) | 415 / 6Max (With Tools) | -3.90 |
General Knowledge
Kimi K3 领先 2/2| 评测项 | Hy4 preview | Kimi K3 | 分差 |
|---|---|---|---|
| CritPt | 16.902 / 3Thinking High (No Tools) | 23.401 / 3最高（无工具） | -6.50 |
| HLE | 55.4015 / 185Thinking High (With Tools) | 5614 / 185Max (With Tools) | -0.60 |
Productivity Knowledge
胶着 2/2| 评测项 | Hy4 preview | Kimi K3 | 分差 |
|---|---|---|---|
| GDPval-AA v2 | 1,6788 / 16Thinking High (With Tools) | 1,6867 / 16Max (With Tools) | -8 |
| Office QA Pro | 66.201 / 3Thinking High (With Tools) | 63.302 / 3Max (With Tools) | +2.90 |
科学与综合推理
Kimi K3 领先 1/1| 评测项 | Hy4 preview | Kimi K3 | 分差 |
|---|---|---|---|
| GPQA Diamond | 92.3024 / 226Thinking High (No Tools) | 93.5013 / 226最高（无工具） | -1.20 |
规格对比
| 字段 | Hy4 preview | Kimi K3 |
|---|---|---|
| 发布机构 | 腾讯AI实验室 | Moonshot AI |
| 发布时间 | 2026-08-28 | 2026-07-16 |
| 模型类型 | 推理大模型 | 推理大模型 |
| 架构 | MoE 架构 | MoE 架构 |
| 参数规模 | 7700亿 | 2.8万亿 |
| 上下文长度 | 1M | 1M |
| 最大输出 | 暂无数据 | 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 | Hy4 preview | Kimi K3 |
|---|---|---|
| 文本输入 | ¥6 / 1M tokens | ¥20 / 1M tokens |
| 文本输出 | ¥18 / 1M tokens | ¥100 / 1M tokens |
| 缓存读取 | ¥0.3 / 1M tokens | ¥2 / 1M tokens |
小结
- Kimi K3在以下类目领先:AI Agent - Tool Usage (3/4)、Coding and Software Engineer (4/4)、Agent Level Benchmark (2/3)、General Knowledge (2/2)、科学与综合推理 (1/1)
- 胶着类目:Productivity Knowledge
16 个共同 benchmark 上，Kimi K3 平均高出 5.91 分。
单项差距最大的 benchmark：Program Bench — Hy4 preview 17.50，Kimi K3 77.80（分差 -60.30）。
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。