Kimi K3vsKimi K2.5
在 1 个同模式 benchmark 中，Kimi K3 整体领先：Kimi K3 领先 1 项，Kimi K2.5 领先 0 项，持平 0 项，平均分差 +503.70。另有 23 项测试模式不同，仅供参考。
Kimi K31 项(100%)(0%)0 项Kimi K2.5
评测分数
按能力类目分组，每组内按分差大小排列；共 1 项。
写作与创意表达
Kimi K3 领先 1/1| 评测项 | Kimi K3 | Kimi K2.5 | 分差 |
|---|---|---|---|
| Creative Writing | 2,0825 / 112Normal (No Tools) | 1,57946 / 112Normal (No Tools) | +503.70 |
测试模式不同的成绩
共 23 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 | Kimi K3 | Kimi K2.5 |
|---|---|---|
| ARC-AGI-1 | 94.50Max (With Tools) | 65.33Thinking (No Tools) |
| Vals Index | 57.81Max (With Tools) | 26.30Thinking (With Tools) |
| MedCode | 48.88Max (With Tools) | 39.32Thinking (With Tools) |
| MedScribe | 87.96Max (With Tools) | 76.44Thinking (With Tools) |
| Vibe Code Bench v1.1 | 84.96Max (With Tools) | 17.54Thinking (With Tools) |
| SimpleBench | 60.70Max (No Tools) | 46.80Thinking (No Tools) |
| LiveBench | 79.19Reported best (effort unspecified) | 69.07Thinking (No Tools) |
| BrowseComp | 91.20Max (With Tools + Internet) | 60.60Thinking (With Tools + Internet) |
| EMB (Excel Modeling) | 66.40Max (With Tools) | 28.47Thinking (With Tools) |
| Finance Agent v2 | 54.40Max (With Tools) | 35.79Thinking (With Tools) |
| HLE | 59.80Max (With Tools) | 50.20Thinking (With Tools) |
| Legal Research Bench | 44.20Max (With Tools) | 15.87Thinking (With Tools) |
| Code Migration | 16.10Max (With Tools) | 6.96Thinking (With Tools) |
| Context Arena | 71.75Max (No Tools) | 53.33Normal (No Tools) |
| Text Arena (Coding) | 1,682Max (No Tools) | 1,431Normal (No Tools) |
| APEX-SWE | 48Max (With Tools) | 30.30Reported best (effort unspecified) |
| CritPt | 23.40Max (No Tools) | 0.60Normal (No Tools) |
| GPQA Diamond | 91.92Thinking High (No Tools) | 87.60Thinking (No Tools) |
| SAGE | 54.26Max (With Tools) | 49.87Thinking (With Tools) |
| τ³-Banking | 41.60Thinking Low (With Tools) | 14.20Thinking (With Tools) |
| MCP-Atlas | 82.30Max (With Tools) | 64.40Normal (With Tools) |
| Toolathlon-Verified | 76.50Max (With Tools) | 33Thinking (With Tools) |
| MMMU-Pro | 83.40Max (With Tools) | 73.10Normal (No Tools) |
规格对比
| 字段 | Kimi K3 | Kimi K2.5 |
|---|---|---|
| 发布机构 | Moonshot AI | Moonshot AI |
| 发布时间 | 2026-07-16 | 2026-01-27 |
| 模型类型 | 推理大模型 | 多模态大模型 |
| 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 | 16K |
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 K2.5 |
|---|---|---|
| 文本输入 | ¥20 / 1M tokens | $0.6 / 1M tokens |
| 文本输出 | ¥100 / 1M tokens | $3 / 1M tokens |
| 缓存读取 | ¥2 / 1M tokens | $0.1 / 1M tokens |
小结
- Kimi K3在以下类目领先:写作与创意表达 (1/1)
1 个同模式 benchmark 上，Kimi K3 平均高出 503.70 分。
单项差距最大的 benchmark：Creative Writing — Kimi K3 2,082，Kimi K2.5 1,579（分差 +503.70）。
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。