Microsoft has introduced a small language AI model it says is cost-effective and can create content for social media using less data than larger models.

The model, called Phi-3-mini, can allegedly outperform models that are twice its size, according to Microsoft’s tests, including in language evaluation, coding, and maths.

According to a research paper released by the company, the new model is capable of running on smart phones and local devices, a step to providing an alternative to the data-laden cloud-powered LLMs which drive up costs.

Phi-3-mini is trained on a smaller data set than the likes of GPT-4, the underlying model of Chat-GPT. It measures 3.8 billion parameters, with capabilities Microsoft says is perfect for smaller businesses unable to run larger models for certain tasks.

Smaller AI models tend to be less expensive to run compared to larger models as they require less data inputs and less energy. This, however, usually comes with a compromise of less capability.

In December, Microsoft released an earlier version of their small AI, Phi-2-mini, which it said performed just as well as Meta’s Llama 2 model.

Phi-3-mini is set to be already available on Microsoft’s Azure AI model catalog, as well as Hugging Face and Ollama.

The capabilities of Phi-3 may suit smaller companies looking for more tailored AI customer applications, since their internal databases tend to be smaller than what is necessary to run all-encompassing AI chatbots like ChatGPT and Google’s Gemini.

Recommended reading

• Report: AI to Replace Almost a Third of Working Hours by 2030
• AI Takeover Ushers ‘Feeling Economy’ Era For Tech Workers
• OpenAI & Meta to Release AI Models that Can ‘Plan’ and ‘Reason’

According to the research paper, while it still does not perform as well as cloud-based model, it out performs other small language models, such as Gemma and Llama-3-In.

It’s main struggle is its range of ‘factual knowledge’ as it does have a smaller data set, but this should not be as much of a limitation for models that do not require a larger data set to function.

Next to join the mini are Phi-3-Small and Phi-3-Medium, both with increased parameters compared to mini.