Microsoft recently unveiled a series of new “open” AI models, highlighted by a particularly advanced model that matches the performance of OpenAI’s widely recognized o3-mini in at least one evaluation measure.
Among the newly introduced AI models, all released under permissive licenses, are Phi 4 mini reasoning, Phi 4 reasoning, and Phi 4 reasoning plus. These latest additions belong to Microsoft’s Phi “small model” line, initially launched to provide foundational capabilities to AI developers working in constrained or edge computing environments.
Phi 4 mini reasoning, featuring around 3.8 billion parameters, was trained on approximately one million synthetic mathematics problems generated by DeepSeek’s R1 model, an advanced reasoning-focused AI developed by a Chinese startup. Designed primarily for educational use and intended for lighter-weight devices or embedded applications, this model aims to offer streamlined tutoring abilities within limited computational environments.
Parameters in AI models serve as rough indicators of their overall problem-solving and reasoning capabilities. Generally, models with larger numbers of parameters tend to deliver superior performance compared to smaller counterparts.
The Phi 4 reasoning model, significantly larger with approximately 14 billion parameters, was trained on curated, high-quality web data and utilized demonstrations from OpenAI’s o3-mini model, further emphasizing mathematics, science, and programming applications.
Phi 4 reasoning plus, meanwhile, represents a specific adaptation of Microsoft’s existing Phi-4 model, optimized explicitly to enhance reasoning ability and deliver improved accuracy for certain tasks. According to Microsoft’s own internal testing, Phi 4 reasoning plus matches R1’s performance—despite the substantial disparity in parameter size (671 billion for R1)—and equals o3-mini on the OmniMath benchmark, a rigorous test for mathematical problem-solving.
Microsoft’s stated goal in deploying these latest Phi models revolves around striking an effective balance between computational size and reasoning prowess. By leveraging techniques such as knowledge distillation, reinforcement learning, and curated training datasets, the company asserts that these new models deliver powerful reasoning capabilities typically associated with larger-scale systems, yet remain sufficiently efficient for resource-constrained or low-latency environments.
All three models—Phi 4 mini reasoning, Phi 4 reasoning, and Phi 4 reasoning plus—are now publicly accessible on Hugging Face’s AI platform, accompanied by comprehensive technical documentation to facilitate practical use and further experimentation by developers.
