Meta Open-Sources Large Concept Model, a Language Model That Predicts Entire Sentences
Briefly

Meta has released the Large Concept Model (LCM), a revolutionary language model that leverages a sentence embedding space instead of token embeddings. LCM, which can handle multilingual summarization tasks, demonstrates superior performance compared to the Llama 3.1 model, particularly in its ability to manage long-form content. Built on the SONAR architecture, LCM supports multiple languages and modalities. Meta views LCM as a move toward enhancing scientific diversity, recognizing that further improvements and scaling are essential for competing with existing large language models.
Meta's open-source LCM operates at the sentence level, outperforming Llama-3.1 in multilingual tasks by using a sentence embedding space independent of language.
With LCM's unique architecture, it leverages the SONAR model's capabilities, allowing it to process text and speech across 200 and 76 languages respectively.
Read at InfoQ
[
|
]