Kimi K3
Moonshot's 2.8T open-weight MoE flagship — the largest open model to date, with native vision, always-on thinking, and #1 frontend coding performance
Kimi K3
Moonshot AI • July 2026
Training Data
Not disclosed
Kimi K3
July 2026
Parameters
2.8 trillion (104B active)
Training Method
Sparse MoE (896 experts, 16 active), Kimi Delta Attention, Attention Residuals, Quantile Balancing
Context Window
1,000,000 tokens
Knowledge Cutoff
Not disclosed
Key Features
Largest Open-Weight Model (2.8T) • Always-On Thinking Mode • Native Vision • Open Weights
Capabilities
Coding: Outstanding
Agentic: Outstanding
Long Context: Outstanding
What's New in This Version
Scales from K2.6's 1T to 2.8T total parameters and 256K to 1M context; leads the Frontend Code Arena and scores 90.4 on 1M-token retrieval, narrowing the gap to Claude Fable 5 and GPT-5.6 Sol
Moonshot's 2.8T open-weight MoE flagship — the largest open model to date, with native vision, always-on thinking, and #1 frontend coding performance
What's New in This Version
Scales from K2.6's 1T to 2.8T total parameters and 256K to 1M context; leads the Frontend Code Arena and scores 90.4 on 1M-token retrieval, narrowing the gap to Claude Fable 5 and GPT-5.6 Sol
Technical Specifications
Key Features
Capabilities
Other Moonshot AI Models
Explore more models from Moonshot AI
Kimi K2.6
Moonshot's open-weight 1T MoE flagship for long-horizon coding and multi-agent orchestration with 300-subagent Agent Swarm system
Kimi K2.5
Moonshot AI's open-source visual agentic model with native multimodal capabilities and multi-agent orchestration
Kimi K2
Moonshot AI's flagship model - #1 open-source LLM on LMSYS Arena with exceptional agentic capabilities
