SOPHIA XT
Live in the browser

A small model that
loads one specialist
per question.

Mnema V1 has 360M parameters. Each skill is a sealed specialist, and a router picks the one that fits your question. You can ask it something now.

Figure 1 Routing one question Illustrative
Question

What is a masked-diffusion scheduler?

arch0.71general0.34diagnosis0.20math0.17code0.27dialog0.19BASE360MONE SPECIALISTLOADED PER QUESTIONarch
The six scores are the cosine similarity between the question and each specialist's capsule. The highest score loads into the base and answers. The questions and the numbers here are made up to show the mechanism.
360Mparameters in the edge base
6sealed specialists
16.7 MiBper specialist, a rank-16 delta
18 of 20queries routed correctly at two specialists
How it is built

The skill is
in a sealed unit.

A specialist ships as one .ncn file. Nothing about it is guessed at load time: the file says what it is for and what it is not for.

Adding a domain means adding a file. The base is not retrained.

One .ncn file rank-16 delta, about 16.7 MiB
  1. Weight delta rank 16

    The part that changes the base's behaviour.

  2. Capability capsule what the router reads

    The vector each question is scored against.

  3. Scope and exclusions declared

    What the specialist is for, and what it is not for.

  4. Validation set bound to the file

    The checks it has to pass travel with it.

  5. Provenance record who made it

    Where the specialist came from.

  1. 01Embed once

    The question is embedded a single time.

  2. 02Score every capsule

    Cosine similarity against each specialist's capsule, in closed form. The scores are visible per query.

  3. 03Attach and answer

    The top scorer is attached to the base, and the base answers with it.

What it taught us

Selection is clean at two,
and breaks at three.

With two specialists the router cleared its 85% gate: 18 of 20 queries across two bases, both misses low-margin. With three or more specialists from similar domains, top-1 accuracy fell to 23.3% and stayed under 47% out to six. We tried centring the capsules, enlarging one corpus and replacing closed form with a learned router. Each made it worse or changed nothing.

We published that as a failure and built the next model around it. XE5 does not select. It blends every specialist on every token and takes its weights from a memory descent. The study that measured the break is XE4-S, and the model it led to is XE5.

2 → 3specialists, the point where closed-form routing fell to 23.3% top-1 on similar-domain capsules.
Figure 2 Top-1 routing accuracy as specialists are added
2 specialists
90%
3 specialists
23.3%
4 to 6 specialists
under 47%
Two specialists: 18 of 20 queries routed correctly, both misses low-margin. Three or more from similar domains: 23.3% at three, and under 47% out to six. The last bar marks that ceiling, not a measured mean.
What this page does not claimThe router figures are from a 20-query check at two specialists and a 30-prompt set at three to six. They show where the mechanism holds and where it stops. They are not a benchmark of answer quality, and Mnema V1 is not offered as one. Figure 1 uses made-up questions and scores to show the mechanism.
Try it

Ask it something,
and watch the route.

The live demo shows the six scores first, then the specialist loading, then the answer as it is made.