Glossaire · IA générative

SLM (Small Language Model)

An SLM, or Small Language Model, is a compact language model with typically 1 to 12 billion parameters, as opposed to large models that line up hundreds of billions. Its reduced size makes it far cheaper to run — on the order of 10 to 30 times less in latency and energy — while remaining amply sufficient for well-scoped tasks: classification, information extraction, summarization, or steering an agent within a defined perimeter. The guiding idea is that a large general-purpose LLM is often oversized for a specific need: on a well-bounded task, a properly chosen or specialized SLM reaches comparable quality for a fraction of the cost. This efficiency opens up new uses: local execution, deployment on constrained devices, low-cost sovereign hosting, and the multiplication of specialized agents. The SLM embodies a pragmatic engineering choice: sizing the model to the real need rather than defaulting to the largest model available.

An SLM, short for Small Language Model, is a compact language model. Where a large LLM lines up hundreds of billions of parameters, the SLM typically has 1 to 12 billion. This compactness is not a limitation to be fixed: it is an engineering choice, dictated by the principle that a model should be sized to the real need rather than called at maximum by default.

Compact does not mean weak

The misleading intuition would be to believe that a smaller model is necessarily worse. In reality, on a well-bounded task — classification, information extraction, summarization, answering within a specific domain — a properly chosen or specialized SLM reaches quality comparable to that of a large general-purpose model. The latter is often oversized for the need: its versatility comes at a high cost and adds nothing more when the perimeter is scoped.

Two levers explain this efficiency. On the one hand, a scoped task draws on only a fraction of a large model's knowledge; the vast remainder becomes a useless cost. On the other hand, an SLM can be fine-tuned on a specific domain — a business jargon, an output format, a document type — to the point of surpassing a large general-purpose model on that limited ground, while remaining infinitely lighter to run.

The economic and energy advantage

It is on cost that the gap becomes spectacular. An SLM consumes on the order of 10 to 30 times less in latency and energy than a large LLM for an equivalent answer. This efficiency changes the game: it makes viable uses where calling a large model would be prohibitive — notably the multiplication of specialized AI agents, each confined to a specific task, run in parallel and at low cost.

Reduced latency matters as much as price. A compact model answers in a fraction of the time of a large LLM, which allows real-time interactions, agent loops chaining many calls, or bulk processing over large volumes of documents. At scale, this difference separates an experimental use from a profitable industrial deployment.

New uses opened up by compactness

Because it fits within a reduced footprint, the SLM can run locally, on constrained devices, without passing through a remote service. This portability directly serves digital sovereignty: a compact model is simpler and cheaper to host in France or the European Union, under complete control of the data.

This is a central lever of our approach to sovereign AI: preferring, when the task allows it, an efficient and controlled SLM over systematic dependence on a large general-purpose model hosted abroad.

FAQ

Frequently asked questions

An SLM, or Small Language Model, is a compact language model with typically 1 to 12 billion parameters, versus hundreds of billions for a large LLM. Its reduced size makes it far cheaper to run while remaining sufficient for well-scoped tasks: classification, extraction, summarization, or steering an agent within a defined perimeter.

The difference comes down to size. An LLM lines up hundreds of billions of parameters and aims for versatility; an SLM has 1 to 12 billion and targets specific tasks. On a well-scoped need, the SLM reaches quality comparable to the large model for a cost 10 to 30 times lower in latency and energy.

Because a large general-purpose LLM is often oversized for a specific need: its versatility costs a lot without adding anything. On a scoped task, a well-chosen SLM matches its quality for a fraction of the cost, with reduced latency. It also unlocks local execution and sovereign hosting.

Yes, markedly. An SLM consumes on the order of 10 to 30 times less in latency and energy than a large LLM for an equivalent answer on a scoped task. This gap makes viable uses where calling a large model would be prohibitive, such as multiplying specialized agents run in parallel at low cost.

There are compact families such as Mistral (Ministral, 3B-8B models), Llama in 1B-8B versions, Gemma, Phi or Qwen in small sizes. Many are open and offered in variants from 1 to 12 billion parameters, often specializable on a specific business domain to surpass a large model on that limited ground.

Yes, it is one of its major strengths. Thanks to its reduced footprint, an SLM runs on a modest server, a workstation, even a constrained device, without passing through a remote service. This portability allows hosting in France or the EU, under complete control of the data.

A self-hosted SLM keeps data on your infrastructure, without sending it to a third-party service outside the EU. It simplifies GDPR compliance and directly serves digital sovereignty: a compact model, controlled hosting cost, full control of processing. It is a pillar of our sovereign AI approach.

As soon as the agent's task is scoped: extraction, classification, summarization, or steering within a defined perimeter. A large general-purpose LLM is often oversized for it. The SLM also enables local execution, deployment on constrained devices and low-cost sovereign hosting — with a far better performance-to-price ratio.

Go further