Local models
We deploy open models on your servers or in an EU cloud, measure them on your real cases and, only when the numbers call for it, fine-tune them for your task. Your data stays in-house and the weights are yours.
What it covers: Open models, On your servers or in the EU, Evaluation on your cases, Fine-tuning with LoRA and QLoRA, Catalan and Spanish, Vision on your images, You keep the weights, Evaluation report
Before training, we measure. Often a good open model with your documents is enough; when it is not, we prove it with numbers and train only what is needed.
From a review of the AI you already run to a model trained for your task. All of them start from an evaluation set built on your cases.
Rescue
You already run Ollama, vLLM or LM Studio, but you don’t know how often it is right or how to improve it. We review it, build an evaluation set from your cases and make the first adjustment.
Deployment
We pick the model that best solves your task, install it on your servers or in an EU cloud and leave it running inside your network, with its licence checked.
Fine-tuning
When the evaluation shows that a model with your documents falls short of the quality, speed or cost you need, we fine-tune it with LoRA or QLoRA on your examples.
Vision
Detection, counting or classification trained on photos from your operation: crops, defects, parts or traffic. Small, fast and able to run offline.
It is the method of Sostingut Lab, our tool for evaluating and training models. Every step leaves numbers you can check.
What the model has to do and when you consider it good, in your words and with a threshold.
A baseline: your current system, a simple rule or the model you already use. Without a baseline there is no improvement to prove.
Real cases with their answer and where it comes from, validated by someone on your team.
If a model scores the answers, we first check on a blind sample that it agrees with a person.
Open models running locally and, if your data allows it, a commercial one; with and without your documents.
Fine-tuning when the comparison shows a gap that your documents do not close.
Results with confidence intervals, a technical sheet in the style of the EU AI Act and a new evaluation every month.
Not to start. Fine-tuning a 7–8 billion parameter model with QLoRA fits on a single 24 GB GPU, which can be rented by the hour in an EU cloud. We only recommend buying hardware when usage justifies it.
It depends on the task. For format, style or classification a few hundred to a few thousand good examples usually do. For the model to learn new knowledge it takes more, and giving it your documents is often better. Quality matters more than quantity.
Often it is, which is why we measure first. Training pays off for stable, repeated tasks, when your documents alone don’t reach the quality, speed or cost you need.
Yes. We compare models with good Catalan, such as Salamandra and ALIA from the Barcelona Supercomputing Center, or Gemma, and evaluate in Catalan on your cases, not on impressions.
That is the best starting point. We measure it as it is, see where it fails and decide whether the fix is in the instructions, the document retrieval or a fine-tune.
Not unless you want it to. We work on your infrastructure or in an EU cloud, and personal data never goes to providers outside the EU.
You keep the process and the evaluation. If you like, we re-evaluate every month and retrain when your data changes or a better base model comes out.
In a 30-minute call we look at the task you want to solve, the data you have and whether a local model is worth it for you.