Microsoft Phi-3: Tiny language model, huge implications for enterprise AI adoption

Microsoft announced the release of Phi-3 today, a powerful 3 billion parameter language model that delivers advanced reasoning capabilities similar to much larger models at a significantly lower cost. Developed by Microsoft Research, the new model will be available on the company’s Azure AI platform, allowing businesses to leverage state-of-the-art natural language processing and reasoning for various applications.

“What matters is that we’re able to have a very tiny model with capabilities—in terms of benchmarks, in terms of experience when you play with the model—that rivals much, much larger models, including getting close to something like a GPT-3.5 level,” Sébastien Bubeck, Vice President of Microsoft generative AI, told VentureBeat. “This is really what this is about. It’s not necessarily the type of progress that we were expecting. I think nobody knew the size that you would need to get capabilities that get close to something like GPT-3.5.”

Phi-3 marks the latest achievement in Microsoft’s efforts to explore the limits of compact language models. Starting with the coding-oriented Phi-1 a year ago and progressing through Phi-1.5 and Phi-2, the Phi series has showcased impressive performance on coding, common sense reasoning, and general natural language benchmarks with models as small as 1-2 billion parameters.

Enabling cost-effective AI for enterprises

“As customers have seen what’s possible, every customer has been running to say, ‘Alright, now I need to do something interesting with this,'” Eric Boyd, Corporate Vice President of Azure AI Platform, told VentureBeat. “On Azure, we’re helping those customers build the generative AI applications they need… We will always have the most capable models out there, really pushing the frontier and showing the edge of what’s possible. But we will also have the best model at each price point along the way.”

With Phi-3, Microsoft has developed a general-purpose 3 billion parameter model that exhibits broad capabilities approaching those of industry-leading models like OpenAI’s GPT-3.5, but at a substantially reduced cost and with the flexibility to run on commodity hardware or even smartphones. This breakthrough in parameter efficiency enables transformative AI use cases for enterprises that were previously cost-prohibitive.

Responsible AI at the forefront

Microsoft developed Phi-3 with its Responsible AI principles in mind from the beginning. The model’s training data was screened for toxicity and biases, and additional safety measures were put in place before release. This enables businesses, particularly those in regulated industries, to confidently harness Phi-3’s capabilities.

From a technical perspective, Phi-3 runs on the ONNX Runtime optimized for NVIDIA GPUs and can be deployed in a distributed manner across multiple GPUs or machines to optimize throughput. The model’s architecture utilizes efficient attention mechanisms and optimized numerical precision to achieve high performance with a relatively small number of parameters.

Empowering businesses with advanced natural language AI

“The beauty is, now that you have this foundational layer in a small model, you can bring in your data and fine-tune this general model, and get amazing performance on narrow verticals,” explained Bubeck. “Even if you pick a narrow domain, you need the general intelligence to be good, even in this vertical domain.”

Microsoft’s launch of Phi-3 and its planned integration into the Azure AI platform represent a significant step forward in making large language model capabilities accessible and cost-effective for businesses of all sizes. As more companies aim to operationalize AI and unlock the value of their unstructured data, purpose-built models like Phi-3 will be essential to achieving that vision.