Microsoft, which has been developing small language models (SLMs) for some time, has announced its most-capable SLM family, Phi-3. SLMs can accomplish some of the same functions as LLMs, but are smaller and trained on less data. That smaller footprint makes them well suited to run in a local environment, which means they’re ideal for smartphones, where in theory they would not even need an Internet connection to run. Microsoft claims the Phi-3 open models can outperform “models of the same size and next size up across a variety of benchmarks that evaluate language, coding and math capabilities.”

The first model in this series to be released, Phi-3-mini, has 3.8 billion parameters and “performs better than models twice its size,” Microsoft explains in a news post.

It will be available in the Microsoft Azure AI Model Catalog, as well as on Hugging Face and Ollama, a framework for running models locally. Phi-3-mini will also be available as a Nvidia NIM microservice via a standard API interface that can be deployed anywhere.

Microsoft says Phi-3-small (7 billion parameters) and Phi-3-medium (14 billion parameters) will be “available shortly” in the Azure AI Model Catalog and other model gardens.

“Parameters refer to how many complex instructions a model can understand,” explains The Verge, noting that Phi-3 performs better than its predecessor, Phi-2, released in December, “and can provide responses close to how a model 10 times bigger than it can.”

Microsoft Azure AI Platform Corporate VP Eric Boyd says Phi-3-mini “is as capable as LLMs like GPT-3.5, ‘just in a smaller form factor,’” according to The Verge.

“Some customers may only need small models, some will need big models and many are going to want to combine both in a variety of ways,” Microsoft AI VP Luis Vargas says in the news post.

“Some of the largest language models today, like Google’s PaLM 2, have hundreds of billions of parameters,” writes Ars Technica, noting that “OpenAI’s GPT-4 is rumored to have over a trillion parameters but spread over eight 220-billion parameter models in a mixture-of-experts configuration” and both models rely on “heavy-duty data center GPUs (and supporting systems) to run properly.”

Microsoft competitors including Google, Anthropic and Meta also have their own small AI models, The Verge points out. “Google’s Gemma 2B and 7B are good for simple chatbots and language-related work,” while Anthropic’s Claude 3 Haiku “can read dense research papers with graphs and summarize them quickly,” and the 8 billion parameter version of Meta’s recently released Llama 3 “may be used for some chatbots and for coding assistance.”