LoRA

Low-Rank Adaptation (arXiv 2106.09685) is the standard parameter-efficient fine-tuning (PEFT) method: freeze the base model and inject small trainable low-rank matrices into its layers instead of updating the original weights — typically ~1% of total parameters, capturing most of the quality of a full fine-tune at a fraction of the cost. Because the base stays frozen, you keep separate lightweight adapters per task and hot-swap them.

QLoRA (arXiv 2305.14314) adds 4-bit quantization of the frozen base on top, bringing large-model fine-tuning within reach of a single consumer GPU — the reason a 3090-class card can adapt a 70B model. Rule of thumb from the Fastino guide: models under 70B fine-tune comfortably via QLoRA on one 80 GB GPU; past that you are in distributed-training territory (sharding, significant complexity).

Choose LoRA when you want efficient, high-quality adaptation on hardware that already holds the model in memory, and full fine-tuning only when the model must absorb a large amount of new knowledge rather than adapt style or behavior.

Real-world shape (@martinbowling, on Cua-S1-4B): LoRA adapters on a frozen Qwen3.5-4B — text + multimodal adapters released separately under Apache-2.0 — then SFT followed by RL where the only reward is task completion (92.9% vs 60.1% on GUI-360 computer use, no accessibility tree). Small enough to run on-prem.

Source: Fastino fine-tuning guide — captured 2026-09-25, clippings/fastino-ai-blog-how-to-fine-tune-open-weights-models-model-selection-data-curation-fine-tuning-strategies-and-ev.md.

Links

• LLM fine-tuning — the umbrella; LoRA is the “how many parameters” axis
• SFT — the usual training signal layered on LoRA adapters