Frequently asked
Questions,
answered honestly.
The questions that come up the most. If yours isn't here, email us — we read every one.
Product
What is MingLLM, in one sentence?
A personal superintelligence that runs locally on your Mac across many surfaces — voice (Jarvis), browser (Tensor), your machine (Rocky), coding (Loom + the Tensor Code CLI) — leaving a receipt for every action and doing nothing without your permission.
What are the surfaces — what's the difference between them?
Jarvis is the voice agent: speak, and it acts across mail, calendar, documents, and macOS apps. Tensor is a Chrome extension that reads across your open tabs and answers with citations pinned to the source. Rocky is the on-Mac agent — double-tap ⌘ to talk, drop a file, hand it a task. Loom is the autonomous coding agent that plans, runs, and repairs long-horizon work (it lives at loom.mingllm.com). They're one system; you reach for whichever surface fits the task.
Is this a chat app?
No. There is no chat window. MingLLM is an agent that lives behind the surfaces you already use — your inbox, your browser, your terminal, your Mac. The voice ear is one shortcut away; the rest is silent until you ask.
What can it actually do today?
Read your inbox and triage it. Answer across open tabs with citations pinned to the source. Move calendar events. Run Shortcuts. Open apps. Read your repo and propose diffs. With Loom, plan and execute multi-step coding work end to end. Every action is shown to you before it fires.
Local-first
Does it really run on my machine?
Yes. The model is on your disk. Inference happens on your Apple Silicon. Your memory is stored in MingLLM's app sandbox. The default state has zero network access — turn the wifi off and MingLLM still works for everything that doesn't reach out.
What about cloud calls?
Off by default. If you want a remote tool — search the web, send mail through your provider, hit a third-party API — you grant that one capability and MingLLM shows you each request before it fires. You can revoke any of those grants from one screen.
Do you train on my data?
No. We have no pipeline that exfiltrates your local memory. The model on your machine is the model we shipped — your data stays on your machine.
Practical
What hardware do I need?
An Apple Silicon Mac (M1 or newer) running macOS 14 (Sonoma) or later, with 16 GB of unified memory recommended. An 8 GB mode is on the roadmap for Q3 2026. Disk: about 8 GB for the base model and your local memory. Intel Macs run, just slower.
When are Windows + Linux coming?
After v1 stabilizes — current target is 2027. We built for macOS first to keep scope tight; the underlying runtime is portable, the system-integration layers are not yet.
How much does it cost?
MingLLM is in invite-only preview today. Pricing for v1 is not finalized — but the local-first design means the unit economics aren't dominated by inference cost, which gives us room to do something fairer than per-token billing.
About us
Who's building this?
A small team led by founder and CEO Yiming Beckmann, building out of Palo Alto, California since 2025. Hiring is open for research, client engineering, and model training — see the team page.
Why MingLLM and not a frontier lab?
Frontier scale is a liability when your product lives on the user's device. We chose the constraint, then built every layer — training, model, runtime, product — to compound on that decision. That's the moat.