LLM fine-tuning, with your data in France.
Fine-tuning adapts an open-source model such as Llama, Qwen, Mistral or Gemma to your data, your vocabulary and your tasks. Run LoRA, QLoRA or full fine-tuning on NVIDIA GPU instances hosted in Europe or on a reserved cluster in Paris and Lyon, then serve the result behind an OpenAI-compatible endpoint.
1 to 8 GPUs per instance · SSH access · Billed per second · No commitment
What is LLM fine-tuning?
Fine-tuning further trains an open-source model such as Llama, Qwen, Mistral, Gemma or DeepSeek on a smaller dataset built for your task. The model learns your format, your terminology and your use case, without training from scratch.
LoRA
Trains small adapters on top of a frozen model. Fast, low memory.
QLoRA
LoRA on a 4-bit quantized model: large models on a single GPU.
Full fine-tuning
Updates all the weights. Most flexible, needs the most GPU memory.
Fine-tuning changes how a model behaves, while RAG changes what it can look up at query time. Many teams combine both.
RAG
Fine-tuning
What changes
The context sent with each request
The model weights
Best for
Fresh or frequently updated facts
Style, format, domain language, specific tasks
Main effort
Indexing and retrieval pipeline
Curated training data and GPU time
When GPU cost occurs
On every request (longer prompts)
During training runs, then serving
What makes fine-tuning hard, and how we help.
Teams that fine-tune open-source LLMs run into the same four constraints.
GPU memory
Enough memory for your method
Full fine-tuning needs far more memory than LoRA or QLoRA. Use Cloud GPU instances with 1 to 8 NVIDIA GPUs, or a reserved GPU Cluster with RTX PRO 6000 Blackwell (96 GB) or B300 (288 GB HBM3e) GPUs.
Cost
Pay only for training time
Fine-tuning runs are bursty. Cloud GPU instances are billed per second with no commitment: start a run, then stop the instance when it ends.
Data residency
Your datasets stay in France
Training data often holds customer records, internal documents or source code. On Arkane Cloud, a Sovereign AI Cloud, Cloud GPU instances are hosted in Europe, and GPU Clusters and Serverless GPU run on our own servers in Paris and Lyon.
Control
Your stack, your weights
SSH into your instance and use the open-source training libraries you already know. The resulting weights are yours to deploy.
The right GPUs for each fine-tuning step.
Full control: you run your own training stack on GPUs you control, then deploy the resulting model on a serverless endpoint. Same console, same account.
Train · per second
Cloud GPU
On-demand NVIDIA GPU instances hosted in Europe, 1 to 8 GPUs, SSH access, billed per second. Run LoRA, QLoRA or full fine-tuning with the open-source tools of your choice.
Best for: most fine-tuning runs, from single-GPU LoRA to multi-GPU jobs
Explore Cloud GPU →Train · reserved
GPU Clusters
Dedicated multi-node clusters with RTX PRO 6000 or B300 GPUs on our own servers in Paris and Lyon, InfiniBand networking, managed Kubernetes or Bare Metal deployment, reserved for 3 months to 5 years.
Best for: large models and recurring fine-tuning programs
Explore GPU Clusters →Serve · per second
Serverless GPU
Deploy your fine-tuned model from a public or private Hugging Face repo behind an OpenAI-compatible endpoint, with scale-to-zero.
Best for: putting your fine-tuned model in production
Explore Serverless GPU →Recommended GPUs
On reserved GPU Clusters, our own servers in Paris and Lyon: RTX PRO 6000 Blackwell (96 GB) for LoRA and QLoRA on small and mid-size models, or multi-GPU jobs on larger ones · B300 (288 GB HBM3e) for full fine-tuning and the largest models. On Cloud GPU: NVIDIA GPU instances hosted in Europe, 1 to 8 GPUs.
Popular base models
Llama, Qwen, Mistral, Gemma, DeepSeek, GPT-OSS and other open-weight models. Compare candidates first on the Models API.
How much GPU memory do I need?
Memory depends on model size and method. Use these indicative figures to size your Cloud GPU instance or your reserved GPU Cluster before you launch a run.
Model size
QLoRA (4-bit)
LoRA (16-bit)
Full fine-tuning
7–8B
~6–10 GB
1 GPU
~16–24 GB
1 GPU
~120–160 GB
2 GPUs
13–14B
~12–20 GB
1 GPU
~32–48 GB
1 GPU
~240–320 GB
4 GPUs
30–34B
~24–40 GB
1 GPU
~64–96 GB
1 to 2 GPUs
~500–700 GB
8 GPUs or a GPU Cluster
70B
~48–64 GB
1 GPU
~160–200 GB
2 to 3 GPUs
~1.1–1.4 TB
A GPU Cluster
Indicative rules of thumb, not benchmarks. They cover weights, gradients and optimizer states at moderate sequence lengths and batch sizes; long contexts and large batches need more memory. QLoRA assumes a 4-bit base model, LoRA a 16-bit base model, and full fine-tuning mixed precision with Adam. GPU counts assume 80 GB-class GPUs. On reserved GPU Clusters, an RTX PRO 6000 Blackwell has 96 GB and a B300 288 GB HBM3e, so fewer GPUs are needed. Multi-GPU jobs split memory with FSDP or DeepSpeed.
Common tools
Unsloth, Axolotl and Hugging Face TRL with PEFT cover LoRA, QLoRA and full fine-tuning. Install them over SSH on your Cloud GPU instance, like any open-source library.
Dataset format
Most tools expect JSONL, one example per line: chat format (a list of system, user and assistant messages) or instruction format (an instruction, an optional input and the expected output).
How to fine-tune an LLM, in four steps.
Baseline
Test candidate base models on your own prompts with the Models API before training anything.
Prepare
Curate and format your dataset, then launch a Cloud GPU instance sized for your method.
Fine-tune
Run LoRA, QLoRA or full fine-tuning over SSH, evaluate, and push the weights to Hugging Face.
Serve
Deploy the fine-tuned model on Serverless GPU, billed per second with scale-to-zero.
Which product for your fine-tuning project?
A quick guide to match each step of a fine-tuning project to the right product.
Workload
Recommended
Why
Quick baseline before fine-tuning
Check whether prompting alone is enough, billed per token
LoRA or QLoRA on a small or mid-size model
A single instance billed per second, stopped when the run ends
Full fine-tuning of a large model
Up to 8 GPUs per instance, or multi-node with InfiniBand
Serving the fine-tuned model
Any vLLM-compatible Hugging Face model, public or private, with scale-to-zero
Fine-tuned model under high, steady traffic
Dedicated capacity and predictable costs
Questions about LLM fine-tuning
What is fine-tuning in AI?
Fine-tuning is the process of further training a pre-trained model on a smaller dataset specific to your task or domain. It adjusts the model's weights so it follows your format, tone and terminology more reliably than prompting alone.
What is the difference between fine-tuning and RAG?
RAG retrieves relevant documents at query time and passes them to the model without changing its weights. Fine-tuning changes the weights themselves. RAG suits fresh or frequently updated knowledge; fine-tuning suits style, format and specialized tasks. The two are often combined.
What is LoRA fine-tuning, and how is QLoRA different?
LoRA (Low-Rank Adaptation) freezes the base model and trains small adapter matrices, so far fewer parameters are updated than in full fine-tuning. QLoRA applies the same approach to a base model loaded in 4-bit precision, which reduces GPU memory further.
Which GPU do I need to fine-tune an LLM?
It depends on model size and method. On Cloud GPU, NVIDIA GPU instances hosted in Europe with 1 to 8 GPUs handle LoRA and QLoRA on small and mid-size models, and multi-GPU instances handle larger ones. On reserved GPU Clusters, RTX PRO 6000 Blackwell (96 GB) suits LoRA and QLoRA on mid-size models, and B300 (288 GB HBM3e) suits full fine-tuning of the largest models.
Does Arkane Cloud offer a managed fine-tuning API?
No. You run your own fine-tuning stack, with the open-source tools of your choice, on Cloud GPU instances over SSH or on a reserved GPU Cluster. Cloud GPU instances are hosted in Europe, and GPU Clusters run on our own servers in Paris and Lyon, France.
How do I deploy my fine-tuned model?
Push the weights to a public or private Hugging Face repository, then deploy the model on Serverless GPU with your Hugging Face token. The model must be vLLM-compatible. You get an OpenAI-compatible endpoint billed per second, with scale-to-zero.
How much does fine-tuning cost?
Cost depends on the GPU model, the number of GPUs and how long your run lasts. Cloud GPU is billed per second with no commitment, so you only pay for the time your run uses. Current rates are on our pricing page.
Other use cases
Fine-tune your models in France, today.
Launch a GPU instance, run your fine-tuning stack and deploy the result on a serverless endpoint, all from one console.
Cloud GPU billed per second · No commitment · Serve on Serverless GPU with scale-to-zero