Our classification
Our models are general-purpose tabular foundation models (in-context learners). As released under Apache-2.0, they are not themselves a high-risk AI system under the EU AI Act. Their use in a downstream system can be high-risk depending on the deployment context (credit decisions, hiring, healthcare diagnostics, etc.). The deployer is responsible for that classification.
What we provide for downstream high-risk deployments
- Model card documenting training data, evaluations, limitations — aligned to AI Act Annex IV technical documentation requirements
- Reproducible evaluation on a locked OpenML benchmark with seed-fixed results
- Calibrated uncertainty outputs — supports human-oversight requirements (Art. 14)
- Conformity-assessment-support documentation available to enterprise customers
- Post-market monitoring hooks via opt-in telemetry in the Python SDK
What we don't do
We do not perform conformity assessment for your specific deployment. That's the deployer's responsibility under the Act. We provide the documentation and technical signals to make your assessment straightforward.
General Purpose AI obligations
Our models fall below the GPAI threshold (10²⁵ FLOPs of training compute). We voluntarily meet the GPAI provider obligations applicable to any general-purpose AI model: technical documentation, copyright-respecting training data policy, training compute disclosure, and a summary of training content. These are linked from each model card on Hugging Face.
Contact
For AI Act questions specific to your deployment: partnerships@zerooneresearch.ai.