On Wednesday, Microsoft introduced a suite of new “open” AI models, among which the most advanced is noted for its competitiveness against OpenAI’s o3-mini on specific benchmarks. This launch is part of Microsoft’s ongoing effort to enhance its AI capabilities and provide robust tools for developers.
The newly released models include Phi 4 mini reasoning, Phi 4 reasoning, and Phi 4 reasoning plus. These models are categorized as “reasoning” models, indicating their enhanced ability to engage in critical fact-checking when addressing complex problems. This initiative expands Microsoft’s Phi “small model” family, initially introduced a year ago, to support AI developers in creating applications that operate efficiently at the edge.
The Phi 4 mini reasoning model was trained on approximately 1 million synthetic math problems, generated using the R1 reasoning model developed by the Chinese AI startup DeepSeek. With around 3.8 billion parameters, this model is specifically designed for educational applications, such as providing “embedded tutoring” on lightweight devices. In the realm of AI models, the number of parameters is a crucial factor, as it often correlates with a model’s problem-solving capabilities; generally, models with more parameters exhibit superior performance.
The Phi 4 reasoning model, boasting 14 billion parameters, was trained on a combination of “high-quality” web data and curated demonstrations sourced from OpenAI’s o3-mini. This model is ideally suited for applications in math, science, and coding, according to Microsoft’s evaluations. Its extensive training data contributes to its specialized functions, making it a valuable tool for developers focused on STEM-related applications.
Additionally, Phi 4 reasoning plus is an adaptation of Microsoft’s previously released Phi-4 model, transformed into a reasoning model aimed at improving accuracy for specific tasks. Microsoft asserts that this model approaches the performance levels of R1, which contains a staggering 671 billion parameters. Internal benchmarking by Microsoft indicates that Phi 4 reasoning plus successfully matches the performance of o3-mini on the OmniMath test, a well-regarded assessment for math skills.
The Phi 4 mini reasoning, Phi 4 reasoning, and Phi 4 reasoning plus models are now accessible on the AI development platform Hugging Face, accompanied by comprehensive technical reports. This availability not only enhances the toolkit for developers but also positions Microsoft as a leader in the increasingly competitive landscape of AI development.
In conclusion, Microsoft’s latest AI models signify a strategic move toward enriching the capabilities available for AI developers, particularly in educational and technical fields. With their advanced reasoning functionalities and robust training data, these models are set to make a significant impact in the AI domain.
