Moonshot AI
Kimi K3
Kimi K3 is an open-weight, natively multimodal mixture-of-experts (MoE) model from Moonshot AI. It has 2.8 trillion total parameters, of which 104 billion are activated per token, and a native context window of 1,048,576 tokens (1M). Moonshot AI describes it as its most capable model, built for long-horizon coding and agentic knowledge work.
- Model ID
- moonshotai/Kimi-K3
Kimi K3 specifications
| Developer | Moonshot AI |
|---|---|
| Architecture | Mixture of experts (MoE) with hybrid Kimi Delta Attention (KDA) and Gated MLA layers |
| Total parameters | 2.8 trillion |
| Activated parameters | 104 billion per token |
| Experts | 896 experts, 16 selected per token, 2 shared |
| Layers | 93 (69 KDA, 24 Gated MLA) |
| Native context length | 1,048,576 tokens (1M) |
| Native input modalities | Text and images |
| Vision encoder | MoonViT-V2 (401M parameters) |
| Released weights | MXFP4 weights with MXFP8 activations (quantization-aware training) |
| Reasoning | Always reasons before answering; low, high and max effort levels |
| License | Kimi K3 License |
Built for
- Long-horizon coding: sustained engineering sessions across large repositories with terminal tools
- Agentic knowledge work, such as deep research
- Native text and image understanding through the MoonViT-V2 vision encoder
On Alchemy Inference
Supported features
- Streaming
- Tool calling
- Structured outputs
- JSON mode
- Reasoning
- Prompt caching
Input
Text and images
API endpoints
Chat completions, Completions, Responses (OpenAI-compatible)
Frequently asked questions about Kimi K3
Kimi K3 is Moonshot AI's open-weight mixture-of-experts model with 2.8 trillion total parameters (104 billion activated per token) and a native 1M-token context window. It natively understands text and images and is built for long-horizon coding and agentic work.
Moonshot AI's model card lists a native context length of 1,048,576 tokens. The context and output limits available on Alchemy Inference are shown in your dashboard.
Yes. Kimi K3 natively accepts text and images, and on Alchemy you can send both.
Yes. According to Moonshot AI, Kimi K3 always reasons before it answers, with low, high and max effort levels. For multi-turn conversations and tool calls, Moonshot AI recommends passing the model's complete previous responses, including reasoning content and tool calls, back unchanged.
Kimi K3 is released under the Kimi K3 License: MIT-style permissions with additional conditions for large Model-as-a-Service providers and for very large commercial products. Read the full license on Hugging Face before building on the open weights.
With Token Based APIs, call Kimi K3 through Alchemy's OpenAI-compatible API with the model ID moonshotai/Kimi-K3 and pay for the tokens you use. For isolated, reserved capacity, request a Dedicated Deployment from your dashboard; the minimum configuration is 8× NVIDIA B300 GPUs. Join the waitlist for early access; once you're in, the in-app quickstart walks you through your first request.
Current pricing and limits for Kimi K3 are listed in your dashboard once you're in.
Alchemy Inference operates with zero data retention (ZDR).
Start building with Kimi K3.
Join the waitlist for early access and up to $1,000 in free credits.