Hugging Face announces SmolLM, a series of new compact language models that outperform well-known products such as Microsoft, Meta, and Alibaba's Qwen in the same category. These models are designed for personal devices and provide cutting-edge AI experiences without compromising performance and privacy. The SmolLM series offers three carefully planned scales: 135 million, 360 million, and 1.7 billion parameter versions, aiming to adapt flexibly to different computing resource requirements. Despite their small size, they demonstrate extraordinary capabilities in benchmark tests for common-sense reasoning and world knowledge mastery. "Disrupting with 'small is beautiful': How SmolLM shakes up the AI industry" Loubna Ben Allal, the Chief Machine Learning Engineer of Hugging Face's SmolLM project, points out that small models tailored for specific tasks can also perform exceptionally well. "Not every task requires a massive base model, just as we wouldn't use a sledgehammer to drive a nail," she metaphorically explains. "Small models designed for specific tasks can efficiently accomplish them." Specifically, SmolLM-135M outperforms Meta's MobileLM-125M with less training data. SmolLM-360M excels among models with fewer than 500 million parameters, including Meta and Qwen's competitors. The flagship model, SmolLM-1.7B, surpasses Microsoft's Phi-1.5, Meta's MobileLM-1.5B, and Qwen2-1.5B in multiple benchmark tests. Hugging Face's comprehensive open-source approach, from data preparation to training, showcases its uniqueness, aligning with its commitment to open-source values and reproducible research. The cornerstone of success: High-quality data preparation SmolLM's outstanding performance is attributed to its meticulously curated training data. These models are built on the Cosmo-Corpus, which combines Cosmopedia v2 (synthetic textbooks and stories), Python-Edu (educational Python samples), and FineWeb-Edu (selected educational web content). "The exceptional performance of SmolLM once again proves the importance of data quality," emphasizes Ben Allal. "We innovatively combine web data with synthetic data to create a high-quality training set, resulting in the currently most outstanding small models." Democratization of AI and privacy protection The release of SmolLM has profound implications for the accessibility and privacy protection of AI. These models can run directly on personal devices such as smartphones and personal computers without relying on cloud computing, reducing costs and minimizing the risk of privacy breaches. Regarding accessibility, Ben Allal states, "Small and efficient models make AI accessible, unlocking new possibilities for everyone while ensuring complete privacy and lower environmental impact." Leandro von Werra, Head of Hugging Face's research team, further explains the practical significance of SmolLM: "These compact models open up a world of infinite possibilities for developers and end-users. From personalized autocomplete to complex user request parsing, SmolLM enables customized AI applications without the need for expensive GPUs or cloud infrastructure. This is an important step towards making AI more widespread and privacy-focused." The release of SmolLM marks a significant turning point in the field of AI. By enhancing AI accessibility and privacy protection, Hugging Face responds positively to public concerns about the environmental impact of AI and data privacy. With the public release of SmolLM models, datasets, and training code, the global AI community and developers will have the opportunity to explore, optimize, and build upon this innovative language model, collectively driving further advancements in AI technology.