How can we help?

A private, memory-first AI assistant for your Mac (and, over the web, from anywhere you allow). It remembers what matters, routes every request to the best model for the job, and can actually do things — run commands, check servers, read your calendar — with your permission.

During the beta, request access and you'll get a signed download plus a short setup. bimi walks you through picking a local model sized to your Mac and connecting any cloud providers you want to use. Nothing is required to leave your device to get started.

No. bimi measures what each connected model can actually do and picks the best one for every request automatically. You can pin a specific model if you want, but you never have to babysit a dropdown. See How routing works.

Yes, for anything a local model can handle. bimi runs an on-device model (llama.cpp) and only reaches for a cloud provider when you've connected one and the task benefits from it — never for a task you've marked on-device.

Both. Type in the composer, or use voice — bimi can recognise who's speaking and respond by voice too. Anything you can click, you can also do by asking.

When an answer has structure — a set of servers, a schedule, a comparison, a subnet plan — bimi renders it as a rich card, diagram, or table inline, not a wall of prose. It's the same content, just readable at a glance.

Yes. You can interrupt a reply, jump in with a new instruction, or pause and resume — nothing you typed is lost.

Longer jobs — a sweep, a multi-step tool run — keep working in the background while you carry on, and show up in a live panel you can open and drill into.

Signals become candidates, the salient ones promote to memory, and memory is revisited between turns to strengthen, decay, or merge. Every fact carries provenance, a confidence score, and a decay curve. The full walkthrough is on How memory works.

Yes — the Memory inspector shows every stored fact, where it came from, and how confident bimi is. You can correct or remove anything.

Forgetting is a soft tombstone: bimi stops using the fact going forward, but the historical record isn't silently rewritten. Ask and it will tell you it was told to forget it.

bimi reconciles them — the newer, higher-confidence, or user-stated fact wins, and it will surface the conflict rather than quietly guessing.

On your device. Memory is local-first by type; sync (when enabled) is a backup channel, and recall never waits on the network. See Privacy posture.

bimi shows the live model in the sidebar, and the console logs every routing decision — the task, the chosen model, its tier, and why. Filter the console to routing to watch it happen.

It classifies the task, looks up which models are measured capable of it, and picks the cheapest that clears the bar — per call, across every provider you've connected. How routing works →

Yes — pin one and bimi will use it, subject to the privacy ceiling (a pin can never send an on-device task to the cloud).

Local models via llama.cpp, plus the major cloud APIs. Connect the ones you have keys for; bimi routes across all of them by best value.

With your permission: run shell commands, SSH into machines, tail logs, check server health, query the network, read your calendar and mail, take notes, and more. Each capability is a tool bimi can call when a request needs it.

For anything sensitive, bimi asks right in the chat — “to do this I'll run <command>; go ahead?” — and you approve by tapping or by saying yes. There's one consent gate, whether you answer by voice or by click.

A small signed companion that lets bimi do privileged local work the sandbox can't. It's version-matched to the app; if it ever gets out of step, bimi offers a one-click reinstall — no download needed.

Only when you've connected a cloud provider and the task isn't marked on-device. Tasks you mark on-device can never leave your Mac — no cloud model is even eligible. There are no third-party telemetry SDKs.

The app logs, relevant state, and a screenshot — the things we need to reproduce the problem. Reports are encrypted to us. You can review what's attached before it sends.

Yes. Memory is yours to inspect and erase, and erase means erased — across every place bimi keeps it.

Yes — bimi runs on Linux through the web interface and a native app, with the same experience as the Mac except for Apple-only tools and sandbox limits. You get the same memory, routing, and tools.

Yes, if your configuration allows it — the same chat, memory, models, and settings, from anywhere you permit.

bimi is in beta now. Pricing is a simple subscription — $14.99/month or $129.99/year, each with a 30-day free trial — plus a free tier; connecting your own model providers means you pay them directly for usage, and bimi keeps that cost visible per model.

Not to get started — the on-device model works without any. To use cloud models, connect your own provider keys; bimi never marks them up and shows you the running cost.

Tap Reinstall & Relaunch on the notice — it copies a fresh, version-matched helper, clears the stale state, and restarts it. No download required.

bimi fails over automatically to another capable model, then another provider, without dropping your turn. If one provider is down, you'll usually still get an answer from another. The console's routing filter shows exactly what happened.

Correct it in the Memory inspector, or just tell bimi the right answer — it will update the fact and supersede the old one. If a confident fact was wrong, that's worth a bug report so we can trace where it came from.

Send a report from inside bimi (below) — it captures the logs and a screenshot automatically, which is the fastest path to a fix.