MoonshotAI iconMoonshotAI: Kimi K2

Model Type

Proprietary model icon

Proprietary Model

API access only

Recommended Use Cases

Text Generation

Try Kimi K2

Kimi K2 is Moonshot AI's open-weight agentic language model released in July 2025, featuring 1 trillion total parameters with 32 billion active, optimized for coding, tool use, and reasoning tasks.

Overview

Kimi K2 is a large-scale open-weight model designed for agentic intelligence. Trained on 15.5 trillion tokens, it excels at coding, reasoning, and tool use while being released under a modified MIT license. On its release day, Kimi K2 had the most downloads on Hugging Face, making it the first Chinese open model to rank among the world's best on LMArena.

Key Features

  • 1 trillion parameters: Massive MoE architecture with 32B active per forward pass
  • 384 experts: More than DeepSeek-V3's 256 experts for improved performance
  • 128K context window: Extended context for complex tasks
  • Open weights: Released under modified MIT license
  • MuonClip optimizer: Novel optimizer for stable large-scale MoE training

Technical Specifications

SpecificationValue
Total Parameters1.04T
Active Parameters32B
Experts384 (selecting 8 per token)
Attention Heads64
Hidden Dimension7168
MoE Hidden Dimension2048
Context Length128K tokens
Training Data15.5T tokens
ArchitectureMoE with MLA (Multi-head Latent Attention)

Role in Series

Moonshot AI's Kimi models offer different capabilities:

  • Kimi K2: Base agentic model, 128K context (this model)
  • Kimi K2 0905: Enhanced coding and frontend, 256K context
  • Kimi K2 Thinking: Extended reasoning with tool orchestration
  • Kimi K2.5: Maximum capability with vision and Agent Swarm

Kimi K2 was replaced by Kimi K2 0905 for most use cases.

Links