Moonshot AI (月之暗面)

Chinese AI company behind Kimi models. Backed by Alibaba. Focuses on developer productivity and agentic capabilities.

Key Products

Kimi K2

  • Architecture: 1T total params, 32B active MoE
  • Context: 128K tokens
  • SWE-bench: 65.8%
  • Native MCP Support: Model Context Protocol
  • Pricing: 0.15 input / 2.50 output per M tokens

Kimi K2.5

  • Appears in OpenClaw rankings
  • Performance tier near M2.1

Model Specifications

SpecKimi K2
Total Parameters1 trillion
Active Parameters32 billion
Experts384 total, 8 selected/token
Training Data15.5T tokens
Context Window128K

Performance Metrics

BenchmarkScore
SWE-bench Verified65.8%
LiveCodeBench53.7%
MATH-50097.4%
Output Speed47.1 tokens/sec
First Token Latency0.53s

Access

  • Platform: platform.moonshot.ai
  • OpenAI-compatible endpoints
  • Free tier available
  • HuggingFace: MoonshotAI/Kimi-K2-Instruct

Notes

  • Founded to focus purely on developer productivity
  • Strong agentic capabilities and native MCP support
  • One of the “Chinese AI Trinity” with Qwen and GLM

[2026-07-20] Kimi K3 发布与市场冲击

来源:Fable 5 × CursorBench / Kimi K3(2026-07-17)。

2026-07-17 发布 Kimi K3——相较 Kimi K2(1T / 128K)大幅升级:

维度Kimi K3
总参数2.8 万亿(2.8T)
上下文100 万(1M)tokens
智能指数(Artificial Analysis)57(接近 GPT-5.5 / Claude Opus 4.8)
单任务成本$0.94(约 Opus 4.8 一半)
前端代码竞技场1679 分(第一)
API 定价3 输入 / 15 输出(每百万 token)
完整权重最晚 2026-07-27 发布

市场冲击:发布当晚催化全球 AI/半导体股抛售(摩根大通称”DeepSeek 2.0”,高盛称”去杠杆事件”)。核心论断:单纯扩大预训练算力不再是获得领先能力的唯一路径,架构/数据质量/RL/后训练效率重要性上升。详见 Kimi K3Scaling Law 之争


Related: Kimi-K2 | Kimi-K3 | China-AI-Model-Landscape-2025-2026 | AI-Models-Landscape-2025-2026