Noema-1 is live

Intelligence,
made legible.

NOETIC is a research lab building Noema — a family of models engineered so that their reasoning can be read, traced and argued with, not merely trusted.

The position

Every frontier model is a black box we agreed to stop opening. We think that trade was premature — and that legibility is not a tax on capability but a source of it.

Read the manifesto

Capabilities

Five things we refuse to hand-wave.

  1. 01 / Reasoning

    It shows its work

    Noema emits a structured derivation alongside every answer. Each step carries the evidence it consumed, so a wrong conclusion can be located rather than merely noticed.

  2. 02 / Attention

    Attention you can read

    Every head is named, grouped and stable across fine-tunes. Our probe tooling maps them to human-legible functions — coreference, arithmetic carry, quotation boundary.

  3. 03 / Memory

    Memory that persists

    A 1.2M-token working context with explicit eviction. Nothing silently falls out of scope; the model tells you what it forgot and why.

  4. 04 / Alignment

    Constraints, not vibes

    Behaviour is specified as auditable predicates checked at decode time. When Noema declines, it cites the clause — and you can read the clause.

  5. 05 / Efficiency

    Small enough to own

    Noema-Micro runs the same interpretability stack in 3.1B parameters, on a laptop, offline, with open weights.

By the numbers

Scale is a means. Here is the means.

Figures describe the Noema-1 production checkpoint as of the latest release. All numbers on this site are illustrative — NOETIC is a fictional lab built as a design exercise.

340BParameters
1.2MContext window
12TTraining tokens
47Published probes

Come look inside.

We are hiring interpretability researchers, systems engineers and one very patient technical writer.