Fine-tune open LLMs on EU GPUs

LoRA and QLoRA fine-tuning on dedicated RTX 3090, 5090 and RTX PRO 6000 instances in Belgium. Per-second billing from $0.08/GPU/hr, root SSH included — and your training data never leaves EU jurisdiction.

  • 24, 32 or 96 GB VRAM per GPU
  • QLoRA from ~13B to 70B-class
  • Per-second billing, from $0.08/hr
  • Training data stays in Belgium
  • Root SSH & JupyterLab

What you can fine-tune an LLM on with a 24 GB cloud GPU

If you want to fine-tune an LLM on a cloud GPU, the first question is not which provider — it is whether your model, method and sequence length fit in VRAM. A dedicated RTX 3090 gives you 24 GB of GDDR6X and roughly 35.6 TFLOPS of FP32 compute. In practice that means: QLoRA (4-bit base weights, trainable low-rank adapters) handles models up to roughly 13B parameters; LoRA on 16-bit weights is comfortable around the 7B class; and full fine-tuning fits only small models, roughly 1–3B. These are rules of thumb — longer contexts and larger batches eat headroom fast.

Just as important is what a 3090 is not for. 70B-class models, multi-node pre-training and H100-scale jobs do not belong on this card, and we will not pretend otherwise — though the 32 GB RTX 5090 and the 96 GB RTX PRO 6000 Blackwell now take QLoRA well past 13B on a single instance. If your target is a Llama- or Mistral-class 7B–13B model adapted to your domain — support transcripts, legal boilerplate, internal documentation — a single 24 GB GPU is usually all the hardware the job needs.

A typical run, start to finish

Create an account on the console, load prepaid credit and launch an instance — it takes minutes, no procurement cycle. SSH in as root (or open the included JupyterLab), install axolotl or unsloth with pip, pull your base model, point the trainer at your dataset and start the run. Checkpoints go to instance storage at $0.15/GB/mo; when training finishes, pull the adapter down at $0.02/GB and stop the instance. Because billing is per second, a short smoke-test epoch to validate your config costs cents, not a rounded-up hour.

Cost math that favours experiments

The arithmetic is deliberately boring. A run that occupies an RTX 3090 for four hours costs $0.48 on its Balanced tier ($0.12/hr) or $0.32 on Flexible ($0.08/hr). How long your run takes depends on your dataset and settings, so time one epoch and multiply — no commitments, and paused time on the grid-aware tiers is never billed. Flexible instances pause first during grid events and resume with processes and GPU memory intact, which makes them a natural fit for checkpointed training that can wait for clean power.

Fine-tuning with proprietary data, under EU law

Fine-tuning is the moment your most sensitive data — the corpus itself — touches someone else's hardware. EponEdge instances are dedicated machines (not a marketplace) in European-owned micro-datacenters in Belgium, so datasets, checkpoints and adapters stay under EU jurisdiction end to end, with no exposure to the US CLOUD Act. For teams that must document processing, see GDPR-compliant GPU hosting; for multi-GPU or reserved capacity, talk to sales.

Fit guide

What fits in 24 GB

Approximate fine-tuning capacity of one RTX 3090 (24 GB). Actual headroom varies with sequence length, batch size and optimizer.
Method Model size Fits on one 3090? Notes
QLoRA (4-bit) Up to ~13B Yes The workhorse setup for Llama/Mistral-class adaptation
LoRA (16-bit) ~7B class Yes Faster iteration; keep batches and context modest
Full fine-tune Small models, ~1–3B Yes Optimizer states dominate memory — small models only
Any method 70B and above No Out of scope for a single 24 GB card — step up to the 96 GB RTX PRO 6000

RTX 3090 tiers: Guaranteed $0.25 · Balanced $0.12 · Flexible $0.08 per GPU/hr, billed per second. RTX 5090 $0.65 / $0.44 / $0.26 · RTX PRO 6000 $1.49 / $0.99 / $0.59. Storage $0.15/GB/mo · bandwidth $0.02/GB. See full pricing details.

FAQ

Fine-tuning on EponEdge, answered

What size LLM can I fine-tune on a 24 GB GPU?

As a rule of thumb: QLoRA handles models up to roughly 13B parameters on 24 GB, LoRA in 16-bit is comfortable around the 7B class, and full fine-tuning fits only small models (roughly 1–3B). Exact headroom depends on sequence length, batch size and optimizer settings.

How much does a fine-tuning run cost on EponEdge?

You pay per second at the tier rate your instance launched with. The arithmetic is simple: a run that occupies an RTX 3090 for four hours costs $0.48 at its Balanced rate of $0.12/hr, or $0.32 on Flexible at $0.08/hr. Run time itself depends on your dataset, sequence length and method, so budget from your own test epoch.

Which pricing tier is best for fine-tuning?

Balanced (from $0.12/hr) suits most attended training runs — it pauses only during severe grid stress, which is rare in practice. Flexible (from $0.08/hr) is ideal for checkpointed jobs that can tolerate pauses; paused time is never billed and instances resume with processes and GPU memory intact. Guaranteed (from $0.25/hr) is for hard deadlines.

Can I fine-tune Llama or Mistral models on EponEdge?

Yes. Open-weight models in the 7B–13B class are the sweet spot for a 24 GB RTX 3090 with LoRA or QLoRA, and the 32 GB RTX 5090 and 96 GB RTX PRO 6000 Blackwell take you to larger models on a single card. You get root SSH on a dedicated instance, so any framework that runs on CUDA — Axolotl, Unsloth, Hugging Face TRL — installs and runs as it would on your own machine.

Does my training data stay in the EU?

Yes. EponEdge instances run in European-owned micro-datacenters in Belgium, and data and compute sit under EU jurisdiction — GDPR-native, with no exposure to foreign cloud acts such as the US CLOUD Act. Datasets, checkpoints and adapters never need to leave Belgium.

Launch a dedicated GPU and start your first fine-tuning run in minutes — or ask us about multi-node clusters and reserved capacity.