A small company wants to classify requests for quotes and extract the service requested, date and missing details. The result is a draft for the person handling email. To choose where to run the model, I would start with that process and the data it needs.
There is no need to send the customer’s full history if the current message is enough. Nor does the draft need to set a price or promise a date. Reducing the scope makes options easier to compare without turning a trial into an infrastructure project.
Three ways to run the same work
Assume a hypothetical scenario with 500 enquiries a month. All three options must produce the same fields and undergo the same human review. I assign none a current price or quality level.
| Option | Where the text is processed | Costs and upkeep I would include |
|---|---|---|
| Managed API | In the contracted service | Usage, integration, review and monitoring changes |
| Open-weight model on a private server | On contracted, configured infrastructure | Server, deployment, updates, backups and support |
| Open-weight model on local equipment | On the company’s equipment, if the workflow sends no data out | Equipment, electricity, operation, security and continuity |
A managed API avoids operating the inference server, but you still maintain the connection and its permissions. I would check what data it receives, how long it retains it and which services are involved. I would not assume that an API is public in the sense of making the data public.
On a private server, you can control the model deployment, but hosting and backups still involve providers and terms. Locally, you need to verify that the equipment can handle the volume and that logs or auxiliary tools do not send information out. If the computer shuts down, the process becomes unavailable unless you have arranged another way to run it.
Available weights do not mean a uniform licence
Being able to download a model lets you consider self-deployment. It is not enough to conclude that every use is permitted. Mistral explains that licences vary across its models; I would check the exact version’s model card and licence before commissioning its deployment.
I would also test that version using real enquiries authorised for that purpose or examples prepared for the trial. A local deployment producing more incorrect fields may need so much review that it stops being worthwhile. Buying equipment does not establish that the output is useful.
In Spain, hosting data nearby does not establish compliance by itself. GDPR Articles 28 and 32 cover processor safeguards and security measures appropriate to the risk. Location is something to review alongside contracts, access and processing, rather than an automatic certificate.
The condition for switching
I would ask for extraction into a provider-independent field format and keep an evaluation sample. Before changing models, I would check correct fields, omissions, review time and cost per accepted enquiry. Adapting the call is only part of it: responses can change too.
With 500 enquiries, I would calculate the full monthly expense and who can sustain it. Combining providers may make sense if distinct tasks justify it, but adds connections and responsibilities. For this small draft, I would prefer to start with one option that meets the requirements and has a documented exit route.
Sources
- Mistral: licences for its open models — Accessed 10 October 2026. Licences vary by model; the specific terms must be checked. This does not establish a deployment’s quality or cost.
- GDPR: Articles 28 and 32 — Accessed 10 October 2026. Regulates processor safeguards and security measures appropriate to the risk; self-hosting alone does not establish compliance in Spain.
If you are deciding where to process enquiries or documents, you can tell me what data needs to leave and who would maintain the system.