SmolLM launches on-device small LLM family (135M/360M/1.7B) with SmolLM-Corpus

AI Impact Summary

SmolLM introduces a family of on-device friendly language models at 135M, 360M, and 1.7B parameters, trained on the SmolLM-Corpus. The post details data curation strategies (Cosmopedia v2, FineWeb-Edu, Stack-Edu-Python) and evaluation against reasoning and common-knowledge benchmarks, highlighting that smaller models can outperform peers in their size class when trained on curated data. For product teams, this implies a feasible path to offline or privacy-preserving LLM deployment using small models, provided they adopt the SmolLM-Corpus data pipeline and perform benchmark-driven validation on target tasks. Deployment planning should consider integrating SmolLM-Corpus components and matching generation styles and prompts to the intended audience, as outlined in the blog.