Models
One architecture, four temperaments. Every member ships the same derivation trace and the same named-head taxonomy — the difference is what we were willing to spend to get there.
Frontier reasoning · production
Our largest and slowest model, and deliberately so. Noema-1 emits a structured derivation next to every answer: each claim carries the tokens it attended to and the heads that carried the weight. It is the model we point at problems where being wrong quietly is unacceptable.
Multimodal · production
Images, diagrams, screenshots and video keyframes enter the same residual stream as text. Vision heads are named on the same taxonomy, so a claim about a chart can be traced back to the pixels that supported it.
On-device · open weights
3.1B parameters with the entire probe stack intact. It runs offline on a laptop, and we publish the weights because an explanation you cannot reproduce is a press release. Most of our own tooling is built against Micro first.
Interpretability stack · preview
Not a chat model — a toolkit. Probe attaches to any Noema checkpoint and exposes head activations, circuit diffs across fine-tunes, and the predicate log that decides refusals. This is the thing the rest of the lab actually uses all day.
Side by side
Scores are on NOETIC's internal legibility suite and are illustrative only — this is a fictional lab and these are invented numbers.
| Model | Parameters | Context | Modalities | Reasoning | Legibility | Availability |
|---|---|---|---|---|---|---|
| Noema-1 | 340B | 1.2M | Text, code | 94.1 | 88.4 | API |
| Noema-Vision | 210B | 640K | Text, image, video | 87.6 | 79.3 | API |
| Noema-Micro | 3.1B | 128K | Text | 61.4 | 82.7 | Open weights |
| Noema-Probe | — | — | Tooling | — | — | Preview |
Access
Noema-Micro weights, the probe stack and every evaluation script. No account, no waitlist, no telemetry.
Free forever
Noema-1 and Noema-Vision over a single endpoint, with traces returned inline rather than behind a second call.
Metered per million tokens
Six months inside the lab with a checkpoint of your own and the right to publish whatever you find, including the parts we dislike.
Apply twice yearly
The playground runs Micro-scale visualisations directly in your browser.