Kimi K3vsKimi K2.6
在 12 个共同 benchmark 中，Kimi K3 整体领先：Kimi K3 领先 12 项，Kimi K2.6 领先 0 项，持平 0 项，平均分差 +44.55。
Kimi K312 项(100%)(0%)0 项Kimi K2.6
评测分数
按能力类目分组，每组内按分差大小排列；共 12 项。
AI Agent - Tool Usage
Kimi K3 领先 4/4| 评测项 | Kimi K3 | Kimi K2.6 | 分差 |
|---|---|---|---|
| Terminal-Bench 2.1 | 88.304 / 49Max (With Tools) | 53.5648 / 49Thinking (No Tools) | +34.74 |
| MCPMark-Verified | 94.501 / 3Max (With Tools) | 72.803 / 3Thinking (With Tools) | +21.70 |
| MCP-Atlas | 84.204 / 41Max (With Tools) | 69.4030 / 41Thinking (With Tools) | +14.80 |
| OSWorld-Verified | 84.802 / 26Max (With Tools) | 73.1015 / 26Thinking (With Tools) | +11.70 |
Coding and Software Engineer
Kimi K3 领先 3/3| 评测项 | Kimi K3 | Kimi K2.6 | 分差 |
|---|---|---|---|
| Program Bench | 77.801 / 7Max (With Tools) | 48.304 / 7Thinking (With Tools) | +29.50 |
| Kimi Code Bench 2.0 | 72.901 / 3Max (With Tools) | 50.903 / 3Thinking (With Tools) | +22 |
| MLS Bench | 48.302 / 5Max (With Tools) | 26.705 / 5Thinking (With Tools) | +21.60 |
AI Agent - Information Search
Kimi K3 领先 1/1| 评测项 | Kimi K3 | Kimi K2.6 | 分差 |
|---|---|---|---|
| BrowseComp | 91.202 / 56Max (With Tools + Internet) | 83.2015 / 56Thinking (With Tools + Internet) | +8 |
General Knowledge
Kimi K3 领先 1/1| 评测项 | Kimi K3 | Kimi K2.6 | 分差 |
|---|---|---|---|
| HLE | 5617 / 190Max (With Tools) | 5422 / 190Thinking (With Tools + Internet) | +2 |
Text Embedding
Kimi K3 领先 1/1| 评测项 | Kimi K3 | Kimi K2.6 | 分差 |
|---|---|---|---|
| Context Arena | 71.7559 / 126Max (No Tools) | 51.8886 / 126Normal (No Tools) | +19.87 |
Writing and Creative Capabilities
Kimi K3 领先 1/1| 评测项 | Kimi K3 | Kimi K2.6 | 分差 |
|---|---|---|---|
| Creative Writing | 2,0712 / 99Normal (No Tools) | 1,72521 / 99Normal (No Tools) | +345.70 |
科学与综合推理
Kimi K3 领先 1/1| 评测项 | Kimi K3 | Kimi K2.6 | 分差 |
|---|---|---|---|
| GPQA Diamond | 93.5016 / 271Max (No Tools) | 90.5042 / 271Thinking (No Tools) | +3 |
规格对比
| 字段 | Kimi K3 | Kimi K2.6 |
|---|---|---|
| 发布机构 | Moonshot AI | Moonshot AI |
| 发布时间 | 2026-07-16 | 2026-04-20 |
| 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 | 2.8万亿 | 1万亿 |
| 上下文长度 | 1M | 256K |
| 最大输出 | 1M | 暂无数据 |
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 | Kimi K3 | Kimi K2.6 |
|---|---|---|
| 文本输入 | ¥20 / 1M tokens | $0.95 / 1M tokens |
| 文本输出 | ¥100 / 1M tokens | $4 / 1M tokens |
| 缓存读取 | ¥2 / 1M tokens | $0.16 / 1M tokens |
| 缓存写入 | 暂无公开价格 | $0.95 / 1M tokens |
小结
- Kimi K3在以下类目领先:AI Agent - Tool Usage (4/4)、Coding and Software Engineer (3/3)、AI Agent - Information Search (1/1)、General Knowledge (1/1)、Text Embedding (1/1)、Writing and Creative Capabilities (1/1)、科学与综合推理 (1/1)
12 个共同 benchmark 上，Kimi K3 平均高出 44.55 分。
单项差距最大的 benchmark：Creative Writing — Kimi K3 2,071，Kimi K2.6 1,725（分差 +345.70）。
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。