Kimi K3
Kimi K3 is a 2.8-trillion-parameter native multimodal agent model with a million-token context. Its architecture combines Kimi Delta Attention, Gated Multi-head Latent Attention, Attention Residuals, and Stable LatentMoE, activating 104B parameters per token. Quantization-aware training uses MXFP4 weights and MXFP8 activations. It supports long-running coding, knowledge-work, and visually grounded agent tasks through Kimi, Kimi Work, Kimi Code, and the official API. Full weights are now available under the custom Kimi K3 License; the initial July 16 release provided hosted access before the weight release.
2026-07-16
2.8T total, 104B activated
Hybrid KDA/Gated-MLA Transformer with Stable LatentMoE, Attention Residuals, and MoonViT-V2
Kimi K3 License
Specifications
- Parameters
- 2.8T total, 104B activated
- Architecture
- Hybrid KDA/Gated-MLA Transformer with Stable LatentMoE, Attention Residuals, and MoonViT-V2
- License
- Kimi K3 License
- Context Window
- 1,048,576 tokens
- Type
- multimodal
- Modalities
- textimagevideo
Benchmark Scores
Advanced Specifications
- Model Family
- Kimi
- API Access
- Available
- Chat Interface
- Available
- Variants
- K3-256K hosted context option
Capabilities & Limitations
- Capabilities
- reasoningcode generationtool usemillion-token contextvisual agentsknowledge work
- Known Limitations
- Thinking history must be preserved correctly; changing models within a session can destabilize generation.The model may act too proactively when user intent is ambiguous.Moonshot recommends supernodes of at least 64 accelerators for serving the full model.The K3-256K service does not accept video; disabling thinking in Kimi Code routes to K2.8 Preview.
- Notable Use Cases
- long-horizon software engineeringresearch agentsvisual productivity workflows
- Function Calling Support
- Yes
- Tool Use Support
- Yes