返回模型列表

Kimi K2 Thinking

moonshotmoonshot/kimi-k2-thinking

API 使用指南

Chat 调用

K2 Thinking 的标准 Chat 调用。模型会在内部生成思考链再输出最终答案。支持流式输出推理 token。

cURL
curl https://api.therouter.ai/v1/chat/completions \
  -H "Authorization: Bearer $THEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "moonshot/kimi-k2-thinking",
    "messages": [
      {"role": "system", "content": "You are a helpful AI assistant."},
      {"role": "user", "content": "Design a distributed caching strategy for a global e-commerce platform."}
    ],
    "max_tokens": 8192,
    "temperature": 0.7
  }'

流式输出 + 推理 token

K2 Thinking 在流式响应中公开推理 token。在请求体中添加 'include_reasoning: true' 即可在最终内容之前收到模型的思维链文本增量。

cURL
curl https://api.therouter.ai/v1/chat/completions \
  -H "Authorization: Bearer $THEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "moonshot/kimi-k2-thinking",
    "messages": [{"role": "user", "content": "Solve 3x^2 + 5x - 2 = 0"}],
    "stream": true,
    "include_reasoning": true
  }'

工具调用(思考+工具)

K2 Thinking 原生地在思维链中将推理 token 与工具调用交错。该模型可在单次推理轨迹中连续进行 200–300 次工具调用并保持状态一致。

cURL
curl https://api.therouter.ai/v1/chat/completions \
  -H "Authorization: Bearer $THEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "moonshot/kimi-k2-thinking",
    "messages": [
      {"role": "user", "content": "Find the latest research papers on MoE architectures and summarise the top 3 findings."}
    ],
    "tools": [{
      "type": "function",
      "function": {
        "name": "web_search",
        "description": "Search the web for information",
        "parameters": {
          "type": "object",
          "properties": {
            "query": {"type": "string", "description": "Search query"}
          },
          "required": ["query"]
        }
      }
    }],
    "tool_choice": "auto"
  }'
事实档案 — 本页每条断言可在此回溯来源
来源URL采集于
发布日期kimi-k2.org ↗2026-05-28已核实
架构huggingface.co ↗2026-05-28已核实
训练成本cnbc.com ↗2026-05-28待核实
磁盘占用(INT4)huggingface.co ↗2026-05-28已核实
许可huggingface.co ↗2026-05-28已核实
支持的推理引擎huggingface.co ↗2026-05-28已核实
训练数据截止——未知
Humanity's Last Exam (w/ tools)lambda.ai ↗2026-05-28已核实
BrowseComplambda.ai ↗2026-05-28已核实
SWE-bench Verifiedlambda.ai ↗2026-05-28已核实
AIME 2025 (w/ Python)www.reddit.com ↗2026-05-28待核实
GPQA-Diamondnist.gov ↗2026-05-28已核实
MMLU-Pronist.gov ↗2026-05-28已核实
CVE-Benchnist.gov ↗2026-05-28已核实
SMT 2025nist.gov ↗2026-05-28已核实
月之暗面发布 Kimi K2 Thinking——支持 200+ 连续工具调用的开源推理模型kimi-k2.org ↗2026-05-28已核实
NIST CAISI 评估 Kimi K2 Thinking——发布时最强的中国开发开源模型nist.gov ↗2026-05-28已核实
Kimi K2 Thinking 在 2× M3 Ultra 上通过 mlx-lm 以 15 tok/s 运行simonwillison.net ↗2026-05-28已核实
Kimi K2(Instruct)和 Kimi K2 Thinking 有什么区别?kimi-k2.org ↗2026-05-28待核实
我能在普通消费级硬件上运行 Kimi K2 Thinking 吗?lambda.ai ↗2026-05-28待核实
K2 Thinking 与 DeepSeek-R1 相比如何?lambda.ai ↗2026-05-28待核实
K2 Thinking 是完全开源的么?huggingface.co ↗2026-05-28待核实
帮助与联系