“The counterparty's markup is back. What moved risk onto our client?”
Reviewed before the call, not after.
Private beta · Windows + NVIDIA RTX
One Lumi.
Many possibilities.
Frontier AI that runs locally, on your own PC. Your files stay on it.
Hand Lumi the contract markup, the patient record, the quarter's ledger. A team of agents reads, cross-checks and researches, then hands back finished work with every source cited.
200+ tok/s · Qwen 35B MoE · RTX 5080 laptop
Agentic mode · confidential work
Built for the files you can't paste into a chatbot.
Switch on Agentic and Lumi plans the work, sends helper agents through your memory, your documents and the web, and shows every step it took. Real screens from the Lumi app, with scripted example files.
“Medication timeline, anything inconsistent, and a de-identified copy for Tuesday's case review.”
The whole pack, read before the patient walks in.
“Reconcile Q3, explain every difference over $500, and draft the note to the client.”
A quarter closed, with last quarter remembered.
“Why did operating margin drop, and is the debt covenant at risk?”
The causes, the chart and the covenant answer, cited to the page.
Confidential
Your files never leave this PC.
Lumi reads them here, on your own graphics card. No upload, no copy on someone else's server. Nothing goes out unless you ask Lumi to fetch something, like a web search or a new model.
Encrypted on your PCEvery chat, file and memory, stored encrypted on your own disk.
Every answer citedPage, clause or message: you can check where each claim came from.
No meterIt runs on hardware you already own. No usage caps, no per-word bill.
Lumi isn't a certified compliance product, and we won't pretend it is. It keeps confidential work on the computer you already control, under the rules you already follow.
LumiPlat · what Lumi does
Everything a desk job asks of it.
Not a chat that tells you what to do. Lumi does it and hands you the document, the spreadsheet or the video.
Reads the fine print
Financial reports, contracts, records. Lumi pulls the tables, compares versions, explains what changed and cites the page.
Writes the deliverable
Fills the form, drafts the document, builds the spreadsheet, redacts and verifies. You review it. You don't retype it.
Remembers every conversation
Encrypted on your PC. Ask what was decided in March and get the answer with a link to where it was said.
Researches with receipts
Reads the web for you, weighs the sources against each other and cites every one.
Codes with a team
LumiCode works in your project with helper agents in the background, and gives way the moment you type.
Edits images and video
Crop, caption, compress. Subtitle a whole video in one sentence.
The voice is all you need.
Speak, interrupt, change your mind mid-sentence. Lumi keeps up, out loud.
Lumi Bare Metal · local inference engine
Blazing-fast agentic inference.
Powering everything.
The engine under every answer. Its scheduler re-optimises itself at every tick, so a graphics card you already own does work that used to need a server rack.
200+
tokens per second
Qwen 35B MoE · RTX 5080 Laptop GPU, 16 GB · Core Ultra 9 275HX · 32 GB RAM
8.8×
faster than Ollama, with 200 agents at once
200 coding agents, every one finished · same models, thinking off · RTX 3080, 10 GB
100
agents on one graphics card
each at its full 16K-token context · 100/100 recall · RTX 3080, 10 GB
~20 ms
for your request to cut in
while 8 background agents run · 20.7–62.6 ms on models that fit on the card · RTX 3080, 10 GB
Proprietary
The Prism Calculus
Our own mathematics for running AI, with a scheduler that re-optimises itself at every tick. It's why one graphics card can carry a hundred agents. We don't publish the maths. We publish the results.
1 GPU · 100 agents · full context
Built by one engineer obsessed with optimisation.
Three hedge funds, where every microsecond counts. Now that obsession runs on your desk.
Already use AI apps? Lumi works with apps built for the OpenAI, Anthropic, Ollama and LM Studio APIs. Point them at Lumi and keep everything local.
200+ tokens/second: Qwen 35B MoE on an RTX 5080 Laptop GPU (16 GB), Intel Core Ultra 9 275HX, 32 GB RAM. 8.8×, 100 agents and ~20 ms: one RTX 3080 (10 GB) with 32 GB RAM, Qwen3 models, October–November 2026; the Ollama comparison used unmodified Aider v0.86.2 with thinking off on both engines. See the full benchmarks.
Zero setup
We handle the engineering.
You enjoy the frontier.
Running frontier AI locally used to mean Python, toolkits, model files and a weekend of settings. Now it means one installer. The rest is our job.
One installer
No Python. No command line. No toolkits. Install it like any other app; the graphics driver you already have is all it needs.
Models, fitted
Lumi downloads each model, checks it and fits it to your graphics card. You never open a config file.
Tunes itself, every tick
Heat, background apps, unplugging the charger: the scheduler adapts as it happens and keeps getting the most from your hardware.
Our mission
Push the boundaries
of local agentic AI.
The most capable computer most people will ever own is already on their desk, idle most of the day. We're making it the best place to run AI: fast, private, and entirely yours.
A whole team on one PC
Lumi for Windows PCs with NVIDIA RTX graphics. The private beta starts here.
More machines
More graphics cards and more kinds of computer.
Questions
Asked, and answered.
Is Lumi an alternative to Ollama?
Yes. Like Ollama, Lumi runs AI models on your own computer. Lumi is a complete app on its own engine, Lumi Bare Metal: one installer, models downloaded and fitted to your graphics card, and agents that work across your documents, memory and the web. Apps built for the Ollama API can point at Lumi. With 200 coding agents running at once on an RTX 3080, Lumi was 8.8× faster than Ollama (same models, thinking off on both). That is a many-agent result, not a claim about every single-user case. Ollama runs on Mac, Linux and Windows; Lumi's beta is Windows with NVIDIA RTX.
How fast is Lumi?
On an RTX 5080 laptop GPU (16 GB), Lumi generates 200+ tokens per second with Qwen 35B MoE. On one RTX 3080 (10 GB), it runs 100 agents at once, each at its full 16K-token context, and your own request cuts in within about 20 ms while 8 background agents work. Speed depends on the model and the graphics card. The full methodology ships with the beta.
Does Lumi work offline?
Yes, for work on your own files. Lumi runs its models on your PC's graphics card, so chat, documents, memory and coding don't need the internet. You need a connection to install Lumi, to download models, and for anything you ask it to fetch, like a web search.
Does Lumi send my data anywhere?
No. Your chats, files and memories are stored encrypted on your own disk and processed on your own graphics card. Nothing goes out unless you ask Lumi to fetch something, like a web search or a new model. When you ask for web research, Lumi sends search queries to the web; it doesn't upload your files.
Can lawyers use Lumi with confidential client documents?
That is the work Lumi is built for. Contracts, markups and bundles are read on your own PC, not uploaded to someone else's server, and every answer cites the clause and page so you can check it. Lumi isn't a certified compliance product and doesn't replace your professional obligations. It keeps confidential work on a computer you already control, under the rules you already follow.
Can doctors, accountants and finance teams use Lumi?
Yes, in the same way. Lumi reads referral packs, ledgers, statements and annual reports on your own PC, cites the page for every finding, and hands back the deliverable: a verified redacted copy, a reconciliation, a client note or a one-page summary. The same caveat applies: Lumi isn't a certified compliance product, so use it within the rules your practice or firm already follows.
What GPU do I need to run Lumi?
A Windows PC with an NVIDIA RTX graphics card. Lumi downloads each model and fits it to your card, so there are no sizes or settings to choose. Our published numbers come from an RTX 3080 (10 GB) and an RTX 5080 laptop GPU (16 GB). Lumi doesn't run on Mac or Linux yet; more graphics cards and kinds of computer are next.
Do I need Python or the command line?
No. Lumi is one installer: no Python, no command line, no toolkits. The graphics driver you already have is all it needs.
Does Lumi work with my existing AI apps and coding tools?
Yes. Lumi works with apps built for the OpenAI, Anthropic, Ollama and LM Studio APIs. Point them at Lumi and the work stays on your PC. For coding, LumiCode works in your project with helper agents in the background and gives way the moment you type.
Is Lumi free?
The core is free for local use. Lumi is in private beta, opening in waves, starting with Windows PCs with NVIDIA RTX graphics.
Private beta
The private beta opens in waves.
Windows PCs with NVIDIA RTX graphics first. The core is free for local use.
One Lumi. Many possibilities.