The episode presented explores Large Concept Models (LCM), a new paradigm for language modeling that focuses on predicting entire sentences as semantic units ("concepts") rather than individual tokens. Using the SONAR embedding space, the LCM approach aims at abstract, multilingual, and multimodal semantic modeling, overcoming the limitations of current Large Language Models (LLM). Diffusion and quantization techniques are employed to enhance the stability and robustness of conceptual representation. Preliminary results demonstrate promising zero-shot generalization capabilities and long-context handling, opening up new prospects for more efficient and cost-effective business applications.