Techniques

LoRA (Low-Rank Adaptation)

A parameter-efficient fine-tuning technique that trains small adapter matrices instead of full model weights.

LoRA freezes the base model's weights and trains small low-rank matrices inserted into attention layers. The result is a fine-tuned model that's 100–1000x cheaper to train and store than full fine-tuning, at ~95% of full-tuning quality on most tasks.

LoRA (and its quantized variant QLoRA) are the default fine-tuning method for open-source LLMs in 2026. You can train a Llama 3.3 70B LoRA on a single A100 in hours; full fine-tuning would need a multi-GPU cluster and days.

Adapters are portable — you ship a small (~100MB) LoRA file that layers onto the base model. This makes per-tenant or per-task customisation practical.

Key points

  • Trains small adapter matrices, not full weights
  • 100–1000x cheaper than full fine-tuning
  • QLoRA adds 4-bit quantization for further efficiency
  • Portable adapters enable per-tenant customisation

Common use cases

Domain adaptation on open-source LLMs
Per-tenant customisation in multi-tenant SaaS
Style / tone / format alignment
Efficient experimentation on multiple fine-tuning ideas

Frequently asked

Related terms

Applying LoRA (Low-Rank Adaptation) in your business?

We help enterprises design, ship, and govern AI systems end-to-end.