Aquin Research

Research

Publications

AMF · Aquin Model Format — A new weight format for internalizing capabilities directly into model weights.

Last UpdatedTitleDescription

Mar 1, 2026AMF EfficiencyFinding better ways to internalize capabilities directly into weights without costing a fortune — more capability in the weights, less compute to get it there.

Mar 1, 2026AMF RobustnessTraining self-awareness directly into the weights so the model can pause, reset, and restart on its own without anything external telling it to.

Mar 1, 2026AMF FactualityA continuously updated knowledge repository retrained into model weights regularly so current and accurate information is absorbed rather than fetched at runtime.

Mar 1, 2026AMF Safety & PrivacySafety and privacy trained so deeply into the weights that unsafe behavior has no representation in the model to begin with — not filtered out after the fact.

Mar 1, 2026AMF Bias & FairnessTraining the model to genuinely not know rather than guess, with a confidence level system built into the weights so honesty and uncertainty are natural responses.

Mar 1, 2026AMF Reasoning & PlanningRunning multiple chains of thought simultaneously inside the weights, approaching a problem from different angles at once so depth of reasoning doesn't cost speed.

Mar 1, 2026AMF Continual LearningWeights structured into tagged checkpoints so new training never overwrites old learning — the model retrieves relevant checkpoints directly rather than searching one massive block.

Mar 1, 2026AMF OCRVisual understanding of objects and text in images trained directly into the weights so the model sees and understands visually the same way it understands language.

Mar 1, 2026AMF DevToolsLetting users inspect and edit what's inside their model in real time — from training data to capabilities inside the weights — the same way a browser's DevTools lets you inspect a live page.

Mar 1, 2026AMF Dynamic TrainingTraining is not a one-shot process. Users can pause a training run at any point and continue it later, picking up exactly where they left off without losing any progress.

Mar 1, 2026AMF Training from ScratchFull control over architecture and initialization, giving users the ability to build and own a model entirely from the ground up on their own data and objectives.

Mar 1, 2026AMF Continuous LearningEvery time the model is called, it trains itself on the incoming data right then, replacing its previous version and continuing forward — always learning, never static.

AMF Efficiency

Finding better ways to internalize capabilities directly into weights without costing a fortune — more capability in the weights, less compute to get it there.

AMF Efficiency

Finding better ways to internalize capabilities directly into weights without costing a fortune — more capability in the weights, less compute to get it there.

Last UpdatedTitle

Mar 1, 2026AMF Efficiency

Mar 1, 2026AMF Robustness

Mar 1, 2026AMF Factuality

Mar 1, 2026AMF Safety & Privacy

Mar 1, 2026AMF Bias & Fairness

Mar 1, 2026AMF Reasoning & Planning

Mar 1, 2026AMF Continual Learning

Mar 1, 2026AMF OCR

Mar 1, 2026AMF DevTools

Mar 1, 2026AMF Dynamic Training

Mar 1, 2026AMF Training from Scratch