Since their emergence, large language models (LLMs) have revolutionized the understanding and automatic processing of natural language, opening the door to advanced applications in industry, customer service, finance, and healthcare. However, their sheer size, computational costs, and data and infrastructure requirements currently limit their relevance for highly specific uses in business.

Why companies are turning to SLMs

Since 2024, a turning point has been observed with the arrival of Small Language Models (SLMs): compact, less expensive to train and use, more economical, and more customizable, they appeal to CIOs and business managers looking for pragmatic, non-generalist AI. According to Gartner, by 2027, SLMs are expected to be three times more widely used than generalist LLMs in B2B environments, as they offer greater accuracy in targeted tasks, enhanced confidentiality, and use fewer computing resources.

SLMs: a response to business requirements

SLMs retain the advantages of LLMs while addressing their limitations: training on business or internal corpora, on-premise or private cloud hosting, and the ability to be integrated directly into business workflows (e.g., extraction of contractual information, industry monitoring, generation of standardized documents). The open-source approach also encourages the emergence of specialized SLMs for sectors such as finance, healthcare, and insurance, where reliability of responses takes precedence over versatility.

Proven benefits and economic impact

Companies that develop or adopt proprietary SLMs see a higher ROI thanks to the detailed exploitation of their business data, fewer contextual errors, and increased security. SLMs are also profitable internally and are sometimes marketed as vertical solutions. The technical paradigm is evolving, with hybrid strategies where an LLM “conductor” delegates complex tasks to expert SLMs.

Outlook and trends for 2025-2027

• Explosion of open source initiatives and “small models” specialized by sector: Llama 3, Mistral Large, Claude 3, adapted BERT, Google Gemma, etc.
• Data-centric strategies: massive use of proprietary data, advanced security, fine-tuning on local corpora.
• Growing integration of the Retrieval-Augmented Generation (RAG) paradigm for contextualized responses in real time.
• Rise of SLMs in responsible digital transformation (better energy efficiency and better data governance).

Useful references to include in the article

• Gartner – « Les SLM, trois fois plus répandus que les LLM en 2027 »
• Techniques de l’Ingénieur – « Les petits modèles de langage rendent l’IA plus accessible »
• Journal du Net – « Les SLM, nouvelle ère pour l’IA d’entreprise »
• MyConnecting-IA – « SLM : l’alternative efficace aux LLM »
• Skimai – « Les 5 meilleurs LLM open source pour l’IA d’entreprise »
• DataDataBoom – “Meilleurs petits modèles de langage en 2025”

This overview, rooted in a neo-realistic and business-oriented vision, enables decision-makers to integrate AI in a responsible, pragmatic, and competitive manner.