Nimi

English

中文

English

中文

Appearance

Local-First Architecture & Multi-Device Execution ​

Nimi can run AI workloads on-device with local models or route requests to cloud providers that the user explicitly configures. "Local-first" does not mean anonymous or whole-product offline operation: Nimi is grounded in account identity, and protected sessions and authorization policies apply in every mode.

Local Inference Execution ​

To run AI tasks locally, prepare the runtime environment and model weights, select the capability's local Loadout, and set the Local route in the app or shared LocalAgent AIConfig. When a job is submitted, Runtime locks that configuration. If local inference fails, Runtime never silently falls back to a cloud provider—keeping privacy, data flow, and API spend predictable. When a Cloud route is selected instead, requests may send relevant context to the configured provider.

Cognition stores each agent's long-term memory locally through Runtime's typed bridge. Memory is enabled per agent, while Knowledge V1 currently has no active ingestion or query interface. Active conversation streams and local audit logs are maintained separately by Runtime.

Offline Operational Boundaries ​

Ongoing local workflows can continue offline as long as the required protected session, capability tokens, local model assets, and security policies remain valid. Cold starts, new sign-ins, fetching fresh Realm character sources, renewing authorization tokens, and calling cloud models require network access. Local caches only optimize performance; they cannot forge committed Realm updates.

Character identity assets, social graphs, economic state, and world definitions (including planned World State/History contracts) belong to Realm. Unsynced local intents are not committed Realm changes. Running locally also does not trigger automatic audit log uploads or unprompted cross-device memory sync.

Multi-Device Topology & Compute Federation ​

Each device runs its own Runtime daemon and hardware-specific loadouts. Signing into the same Realm account across machines shares character identities, but does not automatically sync local hardware loadouts, live conversation streams, or per-agent local memory.

Peer-to-peer compute federation is a future architectural direction, not a current networking layer or installation requirement. Applications must never connect directly to private Runtime debugging ports on remote machines to simulate distributed execution; always use supported protected host interfaces.

Source Basis ​

Documentation for Nimi — the installable, open-source, local-first personal AI product.

Copyright 2026-present Nimi Platform