A hundred agents. One graphics card.

What agentic mode does

Switch on Agentic and Lumi plans the job, then sends helper agents through your memory, your documents and the web. They read, compare, calculate and draft in parallel, and Lumi hands back the finished files with a trail that shows every step it took. In the screen above, one request reconciled 1,284 transactions, explained seven differences, recalled that the same supplier was double-booked last quarter, and delivered a reconciliation and a client note.

Why many agents usually stall on one card

Every agent needs its own working memory for the task in front of it. On one graphics card that memory fills up fast, so most local engines either shorten each agent's context, refuse new work, or make agents wait in line. Your own question waits behind them too. That's why local agents have felt like a demo rather than a tool.

What Lumi measured

• 100 agents at once on one RTX 3080 (10 GB), each at its full 16,384-token context, all 100 answering a recall test correctly.
• 200 coding agents finished 8.8× faster than on Ollama, with every result compiled and tested.
• Your request first: 20.7–62.6 ms for a priority request to cut in while eight agents run, on models that fit on the card.

The method and every model we tested are on the benchmarks page.

A scheduler that tunes itself, every tick

Lumi Bare Metal's scheduler re-optimises itself at every tick, many times a second. It learns how your machine behaves and adapts as conditions change: heat, other apps, unplugging the laptop charger. You never set a batch size or a parallel-request count. The mathematics behind it is ours (we call it the Prism Calculus), and we publish the results rather than the maths.

Coding with a team

LumiCode works in your project with helper agents in the background: exploring, editing and checking while you think. The moment you type, your turn goes first and the helpers give way.

Bring your own agents

Already building with an agent framework? Lumi works with apps built for the OpenAI, Anthropic, Ollama and LM Studio APIs. Point your agents at Lumi and run the whole team on your own PC, with no per-token bill.

Questions