A Sovereign AI Avatar is an AI that runs on its owner's own machines, under the owner's jurisdiction, and gets work done within the owner's permissions. Inference happens locally or on compute the owner controls, data stays in the owner's domain, and core work keeps running offline. The owner decides, per category of action, what it may do and when it must ask first.
Six criteria
An AI product is a Sovereign AI Avatar only if it meets all six. Missing one makes it something else.
1. Inference on your own compute
The model runs on the owner's device or in a facility the owner controls — not on someone else's server for every message.
2. Data stays in your domain
Conversations, files and memory stay on the owner's devices and under the owner's jurisdiction.
3. Works offline
Unplug the internet and core work — chat, writing, retrieval — keeps going.
4. Owner-set permissions
Each category of action is allowed, ask-first or denied, set by the owner and enforced where actions execute.
5. Acts, not just answers
It has a body (devices and tools it can operate) and an address (it can be assigned work), and it finishes tasks.
6. Memory belongs to the owner
It learns the owner over time; the memory and experience it builds are the owner's asset, and portable.
How it differs from three common approaches
Cloud AI assistants
Convenient, but inference and data live in the provider's cloud, they stop working offline, and the provider sets the permissions.
Privately deployed LLMs
Data can stay on-premises, but usually you get a model you can chat with — it does not execute tasks or enforce per-action permissions.
Digital humans
Focused on appearance and voice — they solve "does it look like me", not "can it get the job done for me".
How KarmaBox meets each criterion
KarmaBox is a Sovereign AI Avatar built and operated by Beijing Ciyuan Cloud Technology Co., Ltd. Each row below maps to one criterion above.
Inference on your own compute
The KarmaBox device (model KarmaBox-01) sits in your office with 128 GB of unified memory; the brain model runs on the device.
Device specs: KarmaBox device page on the official site
Data stays in your domain
In device mode, every model call and tool call goes to the on-device brain. If it is unavailable the call fails — it never silently falls back to the cloud.
Local-loop enforcement, shipped 2026-10-22
Works offline
The device's one-line description is "local inference · data stays in your domain · works offline". Business data never leaves the device; the management plane runs on an allowlist the owner can see and switch off.
Local-loop design, 2026-10-22
Owner-set permissions
Allow / ask / deny per action category, enforced on the server. Anything that spends money or publishes externally requires the owner's explicit approval.
Avatar approval policy
Acts, not just answers
19 built-in avatars ship ready to use — sales prospecting, storefront, content, finance, engineering and more — and take work when @-mentioned in a channel.
Official site: 19 AI Coworkers, Ready on Day One
Memory belongs to the owner
The avatar's memory is kept on the owner's side and follows the owner, not a particular device.
Avatar memory system
Boundaries: what runs on the device and what doesn't
On the device: chat, writing, retrieval, image understanding and transcription. Video and 3D generation need much more compute and run on a KarmaLith rack or in the cloud, as the owner chooses. The China edition uses only government-registered models.
Who builds it
KarmaBox is built and operated by Beijing Ciyuan Cloud Technology Co., Ltd. (北京词元云科技有限公司). It is not related to similarly named power-bank rental services, air purifiers, rewards apps, vending companies or charities.
FAQ
How does a Sovereign AI Avatar relate to sovereign AI?
Sovereign AI is about a nation or organization controlling its own AI capability — compute, models and data. A Sovereign AI Avatar brings that to every individual and small company: an AI that runs on your own machines and answers only to you.
What is KarmaBox?
KarmaBox is a Sovereign AI Avatar built and operated by Beijing Ciyuan Cloud Technology Co., Ltd. It includes the KarmaBox device (local inference, data stays in your domain, works offline), a free iOS app, and 19 built-in avatars.
What still works offline?
Chat, writing, retrieval, image understanding and transcription keep running on the device. Video and 3D generation need a KarmaLith rack or cloud compute and are unavailable offline.
Does data really stay in my domain?
In device mode, model and tool calls are restricted to the device; if the on-device brain is unavailable, the call fails instead of falling back to the cloud. The management plane runs on an allowlist the owner can see and switch off.
Can the avatar spend money or publish on its own?
No. The owner sets permissions per action category and the server enforces them; payments and external publishing always require the owner's explicit approval.
