Nano Models

Ellie has reflexes.

About a millisecond. Under 3 MB. Inside Eleplan.

Ellie thinks with a large AI model — it reads, reasons and writes. But not every decision deserves a long think. Is this a question or a task? Is this worth saving? Those calls belong to Nano Models: tiny models Eleplan built and runs itself. They are the reflexes. The large model is the thinking.

Faster than a blink.

Per message. About 2,400× faster than the large-model call it sits beside.

Nano Model — the reflex0.65 ms
Large-model call — the thinking~1.6 s

Medians. 95% of Nano Model decisions finish in under 4 ms.

Smaller than a photo.

Small enough to run on an ordinary CPU inside Eleplan’s own servers.

Eleplan Nano Model< 3 MB
A “small” open AI model (~8 billion parameters)~16 GB at 16-bit

Bars roughly to scale. The Nano Model has about 2 million parameters.

By the numbers.

median time per message
0.65ms
the whole model (1.3 MB compressed)
<3MB
parameters — a small open model has ~8 billion
~2M
retrained, tested and versioned
Weekly

Why it matters.

Instant

The quick calls — what kind of request is this, is this worth saving — finish in about a millisecond.

No outside call

Nano Models run inside Eleplan. The decisions they make need no third-party AI call.

Always on

They run on Eleplan’s own servers — no GPU and no outside service those decisions depend on.

Built for care

When a reflex isn’t sure, it says so and the large model decides. A fast wrong answer is worse than a slow right one.

How they’re made.

  1. 01

    Taught by Ellie

    Ellie’s large model labels thousands of generated, realistic caregiver conversations. That is the textbook.

  2. 02

    One reader, a tiny head per job

    A shared reader understands the words once; a small specialist head makes each decision.

  3. 03

    Must pass an exam

    A fixed test it has to clear before it goes anywhere.

  4. 04

    Shadowed on real conversations

    It runs quietly alongside the real thing and decides nothing until it has earned it.

  5. 05

    Retrained and versioned weekly

    Every week it is retrained, tested again and versioned. When it is unsure, the large model decides.

The line we hold

Anything that changes a health record, or speaks to a family, never runs on a reflex alone.

Those steps go through full reasoning or a deterministic check. Reflexes are for the quick calls around the work, never the work that matters most.

One reader. Many reflexes.

Each reflex is a tiny specialist on top of the shared reader — one per job. Two are running in production today; four are next.

  • Intent32 kinds of requestLive
  • Save detectionis this worth keeping?Live
  • Urgencyhow soon does this matter?Next
  • Calendar relevancedoes this belong on the calendar?Next
  • Document triagewhat kind of document is this?Next
  • Survey answerswhich answer is being given?Next

Measured on production traffic, October 2026.