What LoRA / QLoRA means in practice
Low-Rank Adaptation (LoRA) fine-tunes a model by injecting small trainable matrices into specific transformer layers rather than updating the full parameter set. QLoRA is the quantised variant — base weights frozen at 4-bit precision, adapters trained at 16-bit — which lets a 70B model fine-tune on a single A100. The architectural payoff for enterprise is that one base model can serve dozens of LoRA adapters at once (one per business unit, one per language, one per document type), swappable at inference time. This collapses the model-management problem from "thirty separate fine-tunes" to "one base model and thirty adapter files".
Related terms
Fine-tuning →
Adapting a pretrained model to a specific domain or task by continuing training on a smaller, curated dataset.
Model distillation →
Training a smaller "student" model to imitate the input-output behaviour of a larger "teacher" model.
More in this category
The Sovereign AI Playbook.
Enterprise generative AI that runs on your servers.
- 01Definition. The four non-negotiable properties of a sovereign deployment — and the three patterns that look sovereign but aren't.
- 02Why now. Three forces that converged in 2026: regulator pressure, the open-weights capability gap closing, and unit economics flipping.
- 03Reference architecture. The five-layer stack we deploy: applications → agent runtime → sovereign LLM → RAG + KB → data stores. With model and tooling choices for each.
- 04The regulators. SAMA · RBI · UK ICO · EU AI Act · DPDP · HIPAA — what each requires and how it maps to architecture.
All 62 terms, in plain language
Sovereign AI, RAG, agentic AI, IDP, MLOps and the regulations that shape enterprise AI.