Today, Hugging Face unveiled SmolLM2, a new family of small language models that outperform their larger counterparts in terms of performance while using a lot less computing power. The new models, which are available in three sizes (135M, 360M, and 1.7B parameters) and are published under the Apache 2.0 license, can be deployed on cellphones and other edge devices with constrained memory and processing capability. Most significantly, on a number of important metrics, the 1.7B parameter version performs better than Meta's Llama 1B model.
Hugging Face's model description states that "SmolLM2 demonstrates significant advances over its predecessor, particularly in instruction following, knowledge, reasoning, and mathematics." Using a variety of datasets, including FineWeb-Edu and specialised datasets for mathematics and coding, the largest variation was trained on 11 trillion tokens.
The AI industry is now struggling with the computational needs of running large language models (LLMs), therefore this research comes at a critical moment. The necessity for effective, lightweight AI that can operate locally on devices is becoming more widely acknowledged, even as firms like OpenAI and Anthropic push the envelope with ever-larger models.
The AI industry is now struggling with the computational needs of running large language models (LLMs), therefore this research comes at a critical moment. The necessity for effective, lightweight AI that can operate locally on devices is becoming more widely acknowledged, even as firms like OpenAI and Anthropic push the envelope with ever-larger models.
Many potential users have been left behind by the drive for larger AI models. In order to run these models, costly cloud computing services are needed, which have drawbacks of their own, such as sluggish reaction times, hazards to data privacy, and exorbitant prices that small businesses and independent developers cannot pay.
Many potential users have been left behind by the drive for larger AI models. In order to run these models, costly cloud computing services are needed, which have drawbacks of their own, such as sluggish reaction times, hazards to data privacy, and exorbitant prices that small businesses and independent developers cannot pay.
