This initiative is primarily centered around Microsoft 365 Copilot, the company's artificial intelligence assistant that accompanies its productivity suite. Reports indicate that Microsoft is aiming to incorporate custom and open-source AI models into Copilot, which is currently reliant on OpenAI's technology.
Launched last March, Microsoft 365 Copilot automates routine tasks across the suite's key applications. For instance, the Copilot integrated into Word can condense lengthy documents and create new ones, while the version in Excel offers suggestions for data visualization.
Additionally, the assistant is integrated into various tools used by administrators to oversee their organizations' Microsoft 365 environments. One such tool, Purview, assists in preventing unauthorized use of business data. The Copilot version embedded in Purview provides guidance on tool usage and summarizes alerts related to data leaks.
The report suggests that Microsoft aims to enhance Copilot by integrating new AI models, potentially lowering operational costs for the assistant. Any cost savings could be passed on to customers in the form of reduced prices. Another key focus is to improve the response times of Copilot.
One of the internally developed models that Microsoft may integrate into Copilot is Phi-4, which was introduced earlier this month. This model contains 14 billion parameters, significantly fewer than those found in leading large language models, resulting in lower operational costs.
In internal assessments, Phi-4 demonstrated superior performance compared to a large language model with five times the number of parameters on a benchmark evaluating mathematical capabilities.
Today's report did not identify specific open-source models that may be integrated into Copilot. However, it is likely that Meta Platforms Inc.'s Llama LLMs are among the candidates being considered. This series of LLMs is recognized as one of the most sophisticated within the open-source landscape, often surpassing certain proprietary leading models in various tasks.
The latest version, Llama 3.3, was launched three weeks ago. It delivers output quality comparable to that of a previous Llama model with 405 billion parameters, yet it requires only a fraction of the hardware resources. Additionally, it generally outperformed OpenAI’s GPT-4o in the benchmarks assessed by Meta.
Microsoft is not limiting its efforts to the productivity software market in reducing its dependence on OpenAI. Earlier this year, its GitHub division's GitHub Copilot coding assistant began incorporating LLMs from Google LLC and Anthropic PBC. Nevertheless, Microsoft informed Reuters that "OpenAI remains the company's partner for frontier models."
A Microsoft spokesperson further stated, "We incorporate various models from OpenAI and Microsoft depending on the product and experience.”