Microsoft has just launched Phi-3-Mini, a small but mighty AI model that can carry out tasks like content creation while running locally on smartphones and laptops, making it ideal for companies with limited resources.
Despite being Microsoft’s smallest AI model yet, and being trained on a dataset much smaller than models such as GPT-4, the company claims it can outperform models twice its size because of its training process that prioritizes the quality of data, over quantity.
As more big tech companies like Anthropic and Google try their hand at creating smaller, targeted AI models, Phi-3-Mini is just one of three small models Microsoft is planning to release. We cover what users can expect from the smaller, more affordable model, and how you can test it out for yourself.
Microsoft Launches New AI Model: Phi-3-Mini
For a long time, bigger meant better when it came to artificial intelligence. However, while AI chatbots trained on mammoth datasets like Gemini and ChatGPT have wider applications, many big tech companies appear to be bucking this trend by releasing smaller, targeted models, including Microsoft which just released its smallest AI model yet.
While it might sound like a Sorority house, Phi-3 Mini is one of three compact large language models (LLM) built by Microsoft – with Phi-3-Small and Phi-3-Medium rumored to be released in the coming months. The company’s smallest model has been trained on a data set much smaller than other LLMs, and measuring just 3.8 billion parameters – in comparison with GPT-4’s 1.76 trillion.
🔎 Want to browse the web privately? 🌎 Or appear as if you're in another country?
Get a huge 86% off Surfshark with this special tech.co offer.
However, despite its minimal training data, Microsoft claims the model performs much better than its predecessor Phi-2 – which was released in December – and produces responses up to a standard of models 10 times its size.
“The innovation lies entirely in our dataset for training, a scaled-up version of the one used for phi-2, composed of heavily filtered web data and synthetic data. The model is also further aligned for robustness, safety, and chat format.” – Microsoft
Phi-2-Mini’s success is largely down to its unique training method, which relies heavily on filtered web data and synthetic data, instead of real-world data generated by processes like web crawling. Since synthetic data is much cheaper to source, more diverse, and easier to fine-tune than real-world data, it’s becoming the logical input for many small AI models.
Are Small, Light-Weight AI Models The Future?
Since smaller models like Phi-3 Mini require less computing power, they’re typically cheaper to run and perform better on phones and laptops than larger models. This makes them ideal for companies with smaller AI budgets, that use chatbots for targeted use cases like content generation or solving math problems.
With compact AI models displaying clear advantages over all-purpose ones, Microsoft isn’t the only company focusing on lower-cost targeted chatbots. Google recently introduced two language-focused chatbots, Gemma 2B and 7B, and Anthropic recently launched Claude 3 Haiku – a bot whose primary purpose is to summarize dense research papers.
“If you have a very, very high stakes application, let’s say in a healthcare scenario, then I definitely think that you should go with the frontier model — the best, most capable, most reliable. For other uses, other factors matter more, including speed and cost. That’s where you want to go with Phi-3.” – Microsoft VP Sébastien Bubeck told Axios.
But despite the recent popularity of smaller models, Microsoft isn’t throwing the towels in on larger models just yet. The software manufacturer says that models like Phi-3 aren’t intended to replace large models, but were designed to fulfill goals that models like GPT-4 or Gemini can’t, like running locally on devices.
How Can I Get Microsoft’s Phi-3-Mini AI?
While traditional chatbots aren’t going anywhere, if you’re interested in testing out Microsoft’s new micro-chatbot, Phi-3-Mini is now available for public use.
Microsoft has recently added Phi-3 to its own cloud service platform, Azure model gallery, and the bot is also available on open-source model sites Hugging Face and Olloma. Phi-3-Small, and Phi-3-Medium will also be available on these platforms when they’re released in the future.
If Microsoft’s compact chatbot doesn’t cut it for you, or if you want to compare it to the competition, take a look at our guide to the best AI chatbots in 2024.
