Kimi K3
Language modelOpen weightsCard summarizedDesigned for frontier intelligence across long-horizon coding, knowledge work, and reasoning
Active (latest)Source: Kimi K3 model card (Hugging Face) and tech blog
Use it via API
Where to call it
- Kimi API (Moonshot)
kimi-k3 - Self-host (weights)
moonshotai/Kimi-K3
Supports
- Image input
- Structured outputs
- Streaming
- Reasoning control
- Tool calling
- Web search
- Prompt caching
- Batch API
Good for
- Long-horizon coding with minimal supervision: large repositories, terminal tools, GPU kernels, compilers
- Coding mixed with visual reasoning: frontend, game dev, CAD
- End-to-end knowledge work: deep research, dashboards, documents
Overview
Kimi K3
Designed for frontier intelligence across long-horizon coding, knowledge work, and reasoning
- Cost
- $3.00 in
- Context
- 1M
- Input / output
- $3.00 / $15.00
- Max output
- 1M
- Knows up to
- —
- Released
- Jul 2026
Moonshot's most capable model: a 2.8T-parameter mixture-of-experts model (104B active) built on Kimi Delta Attention and Attention Residuals, with native vision and a 1M-token context. Moonshot calls it the first open 3T-class model and says it still trails Claude Fable 5 and GPT-5.6 Sol overall. model card §1; tech blog
kimi-k3- Input / output per 1M
- $3.00 / $15.00
- Context window
- 1M
- Max output
- 1M
- Knows up to
- —
- Reasoning
- Always on (low / high / max)
- Weights
- Open · Kimi K3 License (modified MIT-style; extra terms for large commercial use)
About this card. A model card is the maker's own account of a model: what it's for, how it was trained and tested, and where it falls short. This page condenses Moonshot AI's Kimi K3 model card (Hugging Face) and tech blog. Every claim carries its source page.