everyone thinks fine-tuning an AI model costs a fortune. tuning Llama 8B on your own GPU costs about a coffee

The short answer

Fine-tuning Llama 3 8B on your own data is cheap and fits a single consumer GPU. Using QLoRA (quantized low-rank adaptation), it needs only ~10–14GB of VRAM — comfortable on a 24GB RTX 3090 or 4090…

QLoRA Llama 8B: ~10–14GB VRAM — fits a single 24GB card (RTX 3090/4090) with headroom.

Cost to rent: ~$0.50 for a single training run on an RTX 4090 at ~$0.44/hr; a full multi-run project is often $10–20.

Time: a few hours for a few epochs on ~10–15k examples.

Own the card and it's free beyond electricity — a used 3090 does this all day.

Full fine-tuning is the expensive one (~60GB+ for 8B, needing datacenter GPUs). QLoRA is what makes it consumer-cheap, at a small quality tradeoff most people never notice.