Small Language Models

Open AI's GPT-4o is a 1 trillion parameter LLM, so does Google's Gemini Pro. In In comparison, Microsoft's Phi-3 is a 3.8B parameter. It is open source, can run locally on devices like cellphone and does not need internet to run. Exactly what a medical, law or many businesses that have to comply with strict data regulations require. Google's latest Pixel phone has Gemini nano, a small language model that does edge computing locally. The small language models can run on 8GM RAM and can generate text at reasonable speed on regular CPU. Last week Apple released OpenELM(open sourced efficient language models) that are small enough to run on a smartphone. With better curated data-set used for pre training and more and more research to enhanced transformer architecture, small language models will only get better.

Tech companies are making smaller AI models to target a wider range of customers, with lower costs and less computing power required. These small models are easier to run on devices and can keep data private. While large language models are still being developed, these smaller ones are seen as a way to get more businesses to adopt AI technology.

Benefits of small models:

• Lower cost and power consumption: Businesses don't need expensive hardware to run AI.
• Privacy focus: Small models can process tasks locally on devices, keeping data internal.
• Accessibility: Enables AI features on mobile devices like smartphones.

Large models still have a place: OpenAI remains committed to developing large models with advanced capabilities like reasoning and planning.

The future of AI: Both large and small models will co-exist, catering to different needs and purposes.