Microsoft has unveiled the Phi-3 series of open small language models (SLMs), heralding them as the most capable and cost-effective in their category. These models, developed through an innovative training method, have surpassed larger models in language, coding, and math benchmarks.

Sonali Yadav, Principal Product Manager for Generative AI at Microsoft, highlighted a shift towards a diverse range of models tailored to specific customer needs rather than a singular focus on large models.

The Phi-3-mini, the first in the series with 3.8 billion parameters, is now accessible through various platforms and outperforms models twice its size. Additional models like Phi-3-small (7B parameters) and Phi-3-medium (14B parameters) are set to follow suit.

Luis Vargas, Microsoft VP of AI, noted the flexibility of using both small and large models depending on the application requirements.

The smaller size of SLMs enables on-device deployment, catering to low-latency AI experiences without the need for network connectivity, with potential applications ranging from smart sensors to privacy-sensitive devices.

While large language models (LLMs) excel in complex reasoning over extensive datasets, SLMs offer an appealing alternative for simpler tasks like query answering and content generation.

Microsoft's breakthrough training technique, inspired by bedtime storybooks, involves curated data sets like 'TinyStories' and 'CodeTextbook' synthesized through a meticulous process of data filtering and generation.

Despite the emphasis on data curation, Microsoft underscores the importance of additional safety measures to address AI risks, applying standard practices for all generative AI models.