Home / Models
Models and model cards
Every model and dataset we have published, in one table: what it is, how big it is, what it was trained on, its licence, whether the weights are public, and its status. Each name links to its card.
- Entries
- 4 models, 2 datasets
- Weights public
- 3 of 4 models
- Updated
- 11 Nov 2026
- Hosted on
- Hugging Face, tetracta
All models and datasets
| Name | What it is | Parameters | Training data | Licence | Weights public? | Status |
|---|---|---|---|---|---|---|
| Z-Next 1B (v4.3) System card | Our own architecture: no attention, no KV cache, a fixed-size state of about 5.2 MiB per conversation. Base model and four tuned variants. | About 947 million | FineWeb-Edu sample and English Wikipedia, 10.0B tokens; a small share of pre-training windows also carried program-generated material that the transformer did not see | Evaluation only | No, card only | Current |
| VAN-1B | Classic transformer, trained as Z-Next's control twin. Base and tuned. | About 0.94B | Same corpus and tokenizer as Z-Next | Apache-2.0 | Yes, base and tuned | Current |
| Chimera-1- | Training-free merge that puts two skills in one model. Failures included. | About 3B | Synthetic task data for the two specialists; the merge itself uses none | Qwen Research License (non-commercial) | Yes | Current |
| llm-xray- | Archived Rational Attention and vanilla checkpoints with seed controls. | 1B class | See the study card | CC-BY-NC-4.0 | Yes | Withdrawn as evidence; archived |
| llm-xray- | Archived Model X-Ray measurements from lesion experiments. | Dataset | None | CC-BY-4.0 | Not applicable | Withdrawn as evidence; archived |
| model-xray- | Archived gallery of Model X-Ray outputs. | Dataset | None | CC-BY-4.0 | Not applicable | Withdrawn as evidence; archived |
Chimera-1 was evaluated by us; it has not been externally reviewed, and an earlier claim that Model X-Ray guided the merge is withdrawn. The Model X-Ray findings that the last three entries once supported were withdrawn on 4 to 6 Sep 2026; see the correction record.
The live village runs at ztown.tetracta.ai. We also run two Hugging Face Spaces: z-town, a static teaser page with video, and model-xray-sample-reports, static sample reports. Neither hosts a model.
What the status tags mean
| Tag | Meaning |
|---|---|
| Current | The card describes the artefact as published, and we stand behind the measured results on it, within the scope stated there. |
| Provisional | Measured, but at small scale, on one run, or before a planned check. The card says which. |
| Withdrawn | We no longer present the claims as evidence. The material stays visible with a link to the correction. |
| In preview | Available to try, but not finished. |
How we write model cards
The standard we follow
Our model cards follow the Hugging Face model card layout and the structure proposed in Mitchell et al. (2019), "Model Cards for Model Reporting". A model that is deployed to users also gets a system card, which describes the deployment: what runs where, which protections are in place and which are not.
What every card contains
- Model details: developer, version, date, type, parameters, licence, whether the weights are available, and a contact.
- Intended use, and uses that are out of scope.
- Training data, including anything synthetic and any gaps in provenance.
- Training procedure at a high level, with compute.
- Evaluation, with confidence intervals where we have them and the caveats of each comparison next to the numbers.
- Limitations, and a list of what is not done yet.
- Risks and safety, including the safety work that was not done.
- Environmental impact in GPU-hours. We do not estimate emissions we did not measure.
- How to access the model, a citation, and a dated changelog.
A number measured on one checkpoint is never attributed to another; readings from intermediate checkpoints are labelled as such. Negative results stay on the card. Where we keep an architecture's mechanism private, the card says so.
How corrections are logged
Every correction gets a dated changelog entry on the affected card that says what changed and why. A withdrawn claim stays visible, tagged Withdrawn, with a link to its correction record. When a Hugging Face card and the card on this site disagree, the site card lists the differences until the Hugging Face card is updated. The Z-Next 1B Hugging Face card was corrected on 10 Nov 2026; the changelog of the card on this site says on which three points. If you find an error, write to [email protected].
Evaluate a model with us
Full benchmarks, confidence intervals and gates written before the run.