Appearance
Qriton HLM is our research line in energy-based models.
Documentation, demos and model releases for Qriton HLM, open to approved teams during controlled early access.
Hopfield layers settle into states that can be inspected.
Most HLM models are built on energy-based (Hopfield) layers. During inference each such layer settles into a stable state. Energy Language can survey those states, edit the weights that shape them and keep a log of every operation.
HLM is a research line and one model option among several in Qriton's work. Explanation and uncertainty signals from energy dynamics are research topics, not product features.
One portal for the whole evaluation.
Approved members work from the same documentation, demos, model cards and release notes.
HLM has model families for language, vision, audio, 3D and edge devices.
Research releases for text, images, audio, 3D and sensor data, each documented with its results and known limits.
- HLM5
- Language models: a transformer trunk with a gated memory for exact-prompt fact edits. HLM5-3B Base is an inference-only research release on Hugging Face.
- HLM3
- Research models used with Energy Language, which inspects their Hopfield layers and makes gated, reversible edits to their weights.
- HLM-Vision, HLM-Spatial, HLM-Audio
- Research models for images, 3D and audio.
- HLM-Micro, HLM-Nano
- Compact research models sized for embedded and edge devices.
HLM5-3B Base and several smaller research checkpoints are public on Hugging Face, each under the licence on its model card. Documentation, demos and the other releases are in controlled early access. Members find the current releases and their limits in the Model Zoo.
Edits to Hopfield weights are checked before they are kept.
Energy Language surveys the basins of a Hopfield layer and edits its weights directly. Edits made through the CLI and BasinSurgeon are kept only if they pass a decision-stability check, and any change can be reverted exactly. So far these edits have been validated on synthetic landscapes; validation on trained HLM3 checkpoints is in progress.
Energy Language is a research tool. A team can survey a layer, propose one edit, see whether the check accepts it, and keep the operation log for review.
Energy Language referenceA person reviews every request.
Access is personal. Research teams, early adopters and pilot projects come first.
01
Request
Tell us your role, your domain and what you want to evaluate.
02
Review
We read every request and reply by email.
03
Invitation
Approved members receive a personal sign-in link. No password.
04
Explore
Documentation, demos, model cards and release notes, in one portal.
Our team can discuss an evaluation with you.
BriefingEarly accessEvaluation on your data