Microsoft has been a key supporter and partner of OpenAI, but it’s clear that the tech giant is not content to let OpenAI dominate the generative AI landscape. In a significant move, Microsoft has introduced a new way to fine-tune its Phi-3 small language model without the need for developers to manage their own servers, and it’s available for free initially.
What is Phi-3?
It is a 3 billion parameter model launched by Microsoft in April. It serves as a low-cost, enterprise-grade option for third-party developers looking to build new applications and software. Despite its smaller size compared to other leading language models, Its performs on par with OpenAI’s GPT-3.5 model. It is designed for coding, common sense reasoning, and general knowledge tasks, making it an affordable and efficient choice for developers.
The Phi-3 Family
The Phi-3 family includes six models with varying parameters and context lengths, ranging from 4,000 to 128,000 tokens per input. Costs range from $0.0003 to $0.0005 per 1,000 input tokens, equating to $0.3 to $0.9 per 1 million tokens. This makes a cost-effective alternative to OpenAI’s GPT-4o mini.
Serverless Fine-Tuning
Microsoft’s new Models-as-a-Service (serverless endpoint) in its Azure AI development platform allows developers to fine-tune Phi-3-small without managing infrastructure. Phi-3-vision, capable of handling imagery inputs, will soon be available via a serverless endpoint as well. For custom-tuned models, Phi-3-mini and Phi-3-medium can be fine-tuned with third-party data.
Benefits and Use Cases
Phi-3 models are ideal for various scenarios, such as learning new skills, improving response quality, and more. For instance, Khan Academy uses a fine-tuned Phi-3 model to benchmark its Khanmigo for Teachers, powered by Microsoft’s Azure OpenAI Service.
Pricing and Competition
Serverless fine-tuning of Phi-3-mini-4k-instruct starts at $0.004 per 1,000 tokens ($4 per 1 million tokens). This positions Microsoft as a strong competitor to OpenAI, which recently offered free fine-tuning of GPT-4o mini for certain users.
