GPU for fine-tuning: LoRA, QLoRA and full training from €0.59 an hour
QLoRA of an 8B model fits on a 16–24 GB card, 32B needs 32–48 GB, and 70B an 80 GB card. LoRA with a 16-bit base needs roughly twice as much, and full fine-tuning of an 8B model already takes a whole 80 GB card. On VRAMfarm training cards start at €0.59 an hour, billed by the minute, and the result can be saved for the next session.
Largest model per method on one card
| GPU | Memory | QLoRA, up to | LoRA 16-bit, up to | Full fine-tune, up to | Price |
|---|---|---|---|---|---|
| RTX 4090 | 24 GB | 8B | 8B | — | €0.59/h |
| RTX 5090 | 32 GB | 32B | 8B | — | €0.76/h |
| RTX A6000 | 48 GB | 32B | 8B | — | €0.72/h |
| L40S | 48 GB | 32B | 8B | — | €1.06/h |
| RTX PRO 6000 | 96 GB | 70B | 32B | 8B | €2.12/h |
| A100 | 80 GB | 70B | 32B | 8B | €1.46/h |
| H100 | 80 GB | 70B | 32B | 8B | €3.90/h |
| H200 | 141 GB | 70B | 32B | 8B | €5.43/h |
| B200 | 192 GB | 70B | 70B | 8B | €16.12/h |
Largest common model size per method on one card, with typical sequence lengths. Full fine-tuning assumes an 8-bit optimiser. More data, longer sequences or bigger batches need more memory. Prices are for one card per hour and change with the market; the price is fixed when you launch. Billing is per minute, minimum 15 minutes.
What people train
LLM on your own data
QLoRA with Unsloth, Axolotl or TRL teaches a model your style, format or domain. A few thousand good examples are usually enough for an 8B model.
Image LoRA
A Flux or SDXL LoRA from 20–30 photos of a person, product or style takes 30–90 minutes on a 24 GB card.
Speech and voice
Fine-tune Whisper on your accent or terminology, or a TTS model on a voice you have the rights to.
Full training and research
Full fine-tuning and pre-training experiments on 80–192 GB cards, or 2–8 cards in one machine with DeepSpeed or FSDP.
Memory per method
QLoRA keeps the base model in 4-bit and trains small adapters: about 1 GB per billion parameters plus activations. LoRA with a 16-bit base needs about 2–3 GB per billion. Full fine-tuning with AdamW needs about 16 bytes per parameter for weights, gradients and optimiser state, so even 8B is around 130 GB; an 8-bit optimiser brings it to one 80 GB card.
Speed versus price
An H100 trains several times faster than an RTX 4090 and costs several times more per hour, so the bill for the same job is often close. Use cheaper 24–48 GB cards for small runs and debugging, then move the final run to an H100 or H200 when time matters.
Do not lose the result
Save checkpoints to /workspace and choose “save and stop”: the folder moves to your next server, even on another card, for €0.03 per GB a month. Without saving, the disk is wiped when the server stops.
What a fine-tune costs
QLoRA of an 8B model on 15 million training tokens with Unsloth on an RTX 4090 takes about 2 hours 50 minutes: roughly €1.65. QLoRA of a 70B model on an H100 for 6 hours: about €23.40. Real time depends on sequence length, batch size and the number of epochs.
Guides on this
12 November 2026 Fine-tune an LLM: how much VRAM and how many GPUs for QLoRA, LoRA and full
QLoRA fits an 8B model on 16 GB and a 70B on one 80 GB card; full fine-tuning needs about 16 bytes per parameter. Memory, cards, tools, hours and cost.
12 November 2026 Train a Flux LoRA on your own photos: 20 images, 24 GB and about two hours
How many photos, which captions, ai-toolkit or FluxGym, 24 or 16 GB, 2,000 steps in 1–2.5 h on an RTX 4090, and what happens to your photos afterwards.
12 November 2026 Multi-GPU training basics: DDP, FSDP and ZeRO on a rented 2–8 GPU machine
When one card isn't enough: data vs model parallel, memory per GPU with FSDP and DeepSpeed ZeRO, NVLink vs PCIe, renting 2–8 cards and avoiding NCCL hangs.
Frequently asked questions
Which GPU do I need to fine-tune Llama 8B?
QLoRA works on a 16–24 GB card, the RTX 4090 is a good choice. LoRA with a 16-bit base needs 24–32 GB. Full fine-tuning needs an 80 GB card with an 8-bit optimiser, or several cards.
Can I fine-tune a 70B model on one GPU?
With QLoRA, yes: 45–60 GB, so one 80 GB card (A100, H100), or 48 GB with Unsloth and short sequences. LoRA with a 16-bit base or full fine-tuning of 70B needs several 80 GB cards.
Can I train on several GPUs at once?
Yes. Pick a machine with 2, 4 or 8 identical cards in the server list and use DeepSpeed, FSDP or Accelerate. All cards are in one machine, so you do not need a network between servers.
What if training runs overnight and the server sits idle after it ends?
We count GPU load as use. About 2 hours after the job ends we email you, and an hour later the server saves your folder and stops, so you do not pay all night.
Start a training run
Top up from €10, pick a card with enough memory and connect over SSH or Jupyter.