root@fr:~$ doctrine --the-right-model
The right model, not the biggest.
Defaulting to the biggest frontier model is almost never the optimal choice. The right model is the one that does the task: often smaller, faster, cheaper, and able to run on French infrastructure. Here is how we right-size.
The benchmark trap
The biggest model on the leaderboard is almost never the right default.
Benchmark-chasing has a simple logic: a new frontier model ships, you wire it in everywhere, you assume it is better at everything. It is an expensive reflex. A huge generalist model pays for its versatility in cost, latency and dependence on a single vendor. On a narrow, repeated task - classify, extract, summarize, route - the extra raw power almost never turns into extra value. You are paying for a supercomputer to sort the mail.
The reflex
the latest model becomes the default on every task, with no check that it actually does them better.
The overhead
versatility billed on every call: more cost, more latency, and a single vendor everything depends on.
The truth
on a scoped task, surplus power does not become surplus value. It becomes an invoice.
Right-sizing
We size the model to the task, not the other way around.
Every task has a profile: volume, sensitivity, latency tolerance, precision requirement. We start from that profile and pick the smallest model that meets the need, not the biggest available. A small specialized model, well placed, often beats a generalist giant on what matters: your cases, your language, your pace.
Volume decides
A task called thousands of times a day is not handled like a rare analysis. Unit cost becomes total cost.
Sensitivity decides
The more sensitive the data, the more sovereign the execution must be. The choice of model is also a choice of jurisdiction.
Latency decides
An agent in an interactive loop needs to respond fast. A small local model often answers before a distant giant has even started.
Precision decides
On a narrow domain, a model specialized on your cases beats a generalist that discovers them on every call.
The smallest model that meets the need. Not the biggest available.
Specialize and distill
A small, well-specialized model beats a big one poorly used.
A small model is not a watered-down version of a big one. With the right techniques, you give it a durable edge on your domain - an edge no generalist holds by default, because it has never seen your cases.
Fine-tuning on your domain
We specialize a smaller model on your documents, your vocabulary and your real cases. It learns your trade instead of improvising it on every call.
Distillation from a big model
We transfer a big model's skill into a smaller one: the small model inherits the know-how, without the invoice or the dependence of the giant.
An asset you own
The specialized model is yours. It runs on your infrastructure, evolves with your business, and does not depend on a foreign vendor's goodwill.
What you gain
Smaller does not mean worse. It means better placed.
Right-sizing changes four dimensions at once, well beyond a marginal saving. And the fourth - sovereignty - is the one a giant frontier model can never give you.
- Cost: a model sized to the task costs a fraction of a generalist giant called every time, at equal volume.
- Latency: a smaller, often local model responds faster. Interactive agents become genuinely usable.
- Privacy: less data leaves the perimeter, less exposure surface. The sensitive stays with you.
- Sovereignty: the best open models run sovereign on French infrastructure, with no US entity ever touching your data.
The conductor
We pick the right model for each task, under the right jurisdiction.
Right-sizing is an ongoing orchestration: each agent has its own model, chosen by task and sensitivity. By default, a sovereign open-weight model hosted in France. For an agent with no sovereignty constraint, bring your own keys (BYOK) and point it at a frontier model. The right model, in the right place, under the right law - decided agent by agent.
| Task | Model | Jurisdiction | Status |
|---|---|---|---|
| Document classification | Small specialized model (open-weight) | France | + Sovereign |
| Structured extraction | Specialized model (fine-tune) | France | + Sovereign |
| Contract analysis | Mistral | France (EU) | + Sovereign |
| Internal brainstorm, non-sensitive | Frontier model (BYOK) | United States | ! Exposed, by choice |
The right model is decided per task, not per account. Which model, where, under which law.
The right reflex
Right-sizing is not a compromise. It is how you make AI both better and sovereign.
We do not trade quality against sovereignty: we get them together. Choosing the right model over the biggest one wins you cost, latency and privacy, and, in the same move, brings execution back onto French infrastructure. The biggest frontier model costs you more and makes you more dependent. The right model does the opposite: it makes you both more capable and more free.