Kimi K3 uses a Mixture-of-Experts architecture. Official materials list 896 experts, 16 selected experts per token, 2 shared experts, and 104B activated parameters out of 2.8T total parameters.
MoE improves efficiency because only part of the model is active for a token, but it does not make the full model small. The system still needs access to the full expert set.
AI handoff prompt and permissions
Copy this prompt into Kimi Code, K3Nova, or another AI agent. Keep approval manual for file writes, shell commands, account actions, and secrets.
Copyable AI prompt
Use this to give the agent the task and safety boundary in one message.
You are my AI agent for this task: explain Kimi K3 as a Mixture-of-Experts model and translate total, active, and expert counts into practical implications.
Start by restating the goal and the permissions you need.
Use official docs or the files I provide before making claims.
Give me a direct answer first, then a short table or checklist.
If commands are needed, show exact copyable commands without a shell prompt.
Ask before writing files, running shell commands, deleting or moving data, logging in, spending money, changing account settings, or handling API keys.
Stop and ask me when a step requires secrets, payment, account access, destructive cleanup, or a permission broader than the task.
Recommended AI permissions
Permission
Give AI
Why
Public research
Allow official docs, cited public pages, and web search results.
Needed to refresh facts without touching private data.
Local files
Do not allow local file access unless you name a folder for inspection.
Most research pages do not need your machine.
Write or shell
Ask before Write, Edit, Bash, or any command execution.
Keeps the task reviewable.
Sensitive actions
Deny login, payment, API keys, account changes, and private screenshots.
Those actions need a human decision.
How to read the facts
Sparse routing
Each token is routed through selected experts rather than every expert, which changes compute cost per token.
Shared experts
Shared experts provide common capacity alongside routed experts.
Hardware implication
The full expert set still has to be stored, so deployment planning remains memory-heavy.
Quick scan cards
Total parameters
2.8T Full model scale
Activated parameters
104B Approximate active portion per token
Experts
896 Specialized parameter groups
Selected experts per token
16 Experts routed for each token
Context length
1,048,576 tokens Maximum listed context window
FAQ
What does MoE mean for Kimi K3?
It means Kimi K3 routes each token through selected experts instead of activating the entire model.
Does MoE reduce storage needs?
No. It reduces active compute per token, but the full expert set still needs to be stored.
Why does MoE matter for self-hosting?
Routing across experts can make interconnect and serving software as important as raw VRAM.
K3Nova