01 · Allonia specialised models
Trained for a specific task
Understanding a customer’s request, choosing the right filters in a catalogue, writing a reply grounded in your data, answering from an FAQ with the sources cited.
Models
Specialised models, hosted in France.
Models trained for customer service and retail, open models on demand and your own models. All served through the same API, in our cloud or in yours.
Our catalogue
Three families, one API.
Use a ready-made model or have your own trained: the way you call it stays the same.
02 · Open models on demand
The best open models, with no infrastructure to run
Mistral, Llama, Gemma and other open models, served with a simple API key. No platform to deploy, no GPUs to manage.
03 · Your own models
Trained on your data, served like the rest
Trained or fine-tuned with NEXUS, then exposed through the same API, with their documentation, metrics and version history.
Why specialised models
More accurate, leaner, safer.
A large general-purpose model can do a bit of everything. For a task that comes up thousands of times a day, a model trained for it does the job better, faster and for less.
-
Domain accuracy
The model has learnt your vocabulary and your tricky cases: fewer instructions to write, fewer errors to correct.
-
Costs under control
Specialised models are more compact, so they respond quickly and use less compute on every request.
-
Confidentiality
Hosted in France or in your own infrastructure: your requests stay within boundaries you know.
-
Controlled releases
Every new version is evaluated, documented and announced before it goes live.
A standard API
Connect in a few lines.
The API follows the OpenAI format: your existing tools and libraries work without a rewrite.
- Chat, completions and embeddings
- One key per application, with its own quotas
- Call logs and usage tracking
- The same API in our cloud or in yours
POST /v1/chat/completions
Authorization: Bearer $ALLONIA_API_KEY
{
"model": "customer-service-model",
"messages": [
{ "role": "user",
"content": "Where is my order 4127?" }
]
}Going live
Measure before you switch.
A model only replaces the one you use today once it has proved itself on your own conversations.
Test set
Real conversations, anonymised and taken from your history.
Blind comparison
Answers from the new model and the current one are judged without knowing which is which.
Corrections
Every discrepancy is analysed and fixed before go-live.
Monitoring in production
Quality, response times and costs are tracked over time.
Frequently asked questions
Our models in practice.
Which open models do you offer?
The open model families most widely used in business, such as Mistral, Llama and Gemma, in versions chosen for the quality of their French. The list changes with the market: we will tell you which models are available when your project starts.
Where are the models hosted?
In data centres in France, or in your own infrastructure if your project requires it. Either way, you know where your requests are processed.
Is our data used to train your models?
No. Your requests are used to answer you. Training on your data only happens as part of a project you decide on, and the resulting model is reserved for you.
Can we use the API without NEO or NEXUS?
Yes. Model access is a product in its own right: an API key is all you need, with no platform to deploy and no need to duplicate your technical teams’ tools.
Can we host the models ourselves?
Depending on the model and your hardware, yes: we work with you on installing them in your private cloud or on your GPU servers.
Demo, API access or workshop
Let’s talk about your project.
A specific question?
A NEO demo, access to the models API, a workshop on your data: tell us what you are looking for and an expert will get back to you.
sales@pertimm.com