On your device, not in a datacenter.
Weights, memory, and compute all sit on your Mac. Nothing leaves unless you explicitly send it. There is no cloud account behind the curtain.
MingLLM is a personal superintelligence that lives on your Mac. Voice is the front door. The browser, your inbox, your calendar, and the documents you already have open are the rooms it walks through. There is no cloud account behind it. There is no telemetry pipe running back to us. The model is on your machine, and the memory it keeps is yours.
We started this because we wanted an assistant that didn't require us to mail our lives to a stranger to use. What follows is the story of how it works — six movements, one shape.
Voice in. Action out. No cloud, no tenants, no telemetry — a personal superintelligence that lives on your machine and waits for the moment you need it. The first AI that doesn't phone home.
Say the thing. MingLLM moves — mail, calendar, the documents open in front of you. Voice sits at the front; a real agent works behind it. Words become work without the dance of pasting and re-explaining.
Across every tab, every page, every paper. Parallel attention held in a single place — so answers are grounded in what's actually in front of you, not in a frozen snapshot of last year's web.
Not just what you said. MingLLM keeps the shape of your week, your project, the way you phrase things — and lets that context bend every answer toward the version that lands for you, not for a stranger.
No silent background automation. Each action MingLLM takes leaves a legible trail — what it read, what it changed, why it chose to. Reviewable, reversible, yours to check before, during, and after.
The model runs on your device. Memory stays on your device. What reaches the network reaches it because you allowed it — not because we defaulted you into it. The intelligence belongs to the person it works for.
01 — Movement
Six movements. One shape.
Weights, memory, and compute all sit on your Mac. Nothing leaves unless you explicitly send it. There is no cloud account behind the curtain.
MingLLM is built around speaking — interrupt-tolerant, low-latency, and aware of who you are. Typing is the fallback, not the design center.
Your projects, your phrasing, your week. MingLLM holds context that survives between sessions and bends every answer toward the version you actually need.
What it read. What it changed. Why it chose to. Each step is a receipt — reviewable while it happens, reversible after the fact, never silent.
The most useful AI tools today require you to mail your private life to a stranger. Your messages, your code, your calendar, your half-formed ideas. We thought the right shape was simpler: the model belongs to the person it works for.
MingLLM is built by a small team in San Francisco. We come from ML research, systems software, and macOS internals. We're making the assistant we wanted ourselves.
2025 · Spring
A handful of researchers tired of pasting their lives into someone else's model started prototyping a local agent that could actually act inside a Mac.
2025 · Summer
Ming (明) means bright, clear, intelligent — the sun and moon together. The naming choice was the first design decision: clarity, not cleverness.
2025 · Autumn
Sub-300ms streaming voice on-device. The first time you could finish a sentence and watch the calendar fill itself in before you took a breath.
2026 · Now
Running daily on the team's machines. Mail, calendar, browser, files. We are letting a small circle of friends in next.
Bright. Clear. Intelligent. Private. Ours.
An intelligence that doesn't ask you to give yourself away to use it.
— The MingLLM team
What comes next
Voice, browser, research, proof. MingLLM shows up where you work — one intelligence, many ways in. Jarvis on the desktop. Tensor in the browser. Receipts everywhere it acts.