Text Gen · Model Guide · 5 min read

Kimi K3

Moonshot's flagship trillion-parameter MoE with million-token context — frontier knowledge at agentic speed.

Basic Information

Kimi K3 is Moonshot AI's frontier text model: a mixture-of-experts architecture with roughly 2.8 trillion total parameters of which about 104 billion are active per token, plus a hybrid attention design that extends the usable context window to around one million tokens. The large expert pool gives it deep world knowledge and strong reasoning, while the small active set keeps generation responsive for its class.

It is a multimodal-family model (K3's open release includes vision components) and shines at agentic workflows: reading entire codebases, long document chains, tool use, and multi-step reasoning. Released with open weights, it is one of the most capable models anyone can self-host — provided the hardware is datacenter-class.

Best use cases

• Agentic coding and repository-scale refactoring (1M-token context swallows whole codebases)
• Deep document analysis: contracts, research corpora, long report chains
• Complex reasoning and planning where smaller models lose the thread
• Private frontier-quality deployments for teams that cannot send data to cloud APIs

Computational Requirements

Kimi K3 is datacenter-class. Total parameters dominate the memory bill regardless of the small active set — the whole model must sit in VRAM.

• Practical for enterprises with multi-node GPU clusters or high-end cloud GPU rentals; not runnable on consumer hardware at any quantization.
• For most individuals, the realistic path is a hosted API (the pricing is very aggressive for the capability class).

Commercial license

Open weights under Moonshot's K-series licensing (Modified-MIT lineage), which permits commercial use and self-hosting. Read the model card before shipping — acceptable-use terms apply, and exact wording can change between releases.