On Wednesday, Microsoft unveiled a new suite of “open” AI models, marking a significant addition to its Phi family. The most advanced model, Phi 4 reasoning plus, demonstrates competitive capabilities with OpenAI’s o3-mini across various benchmarks. These newly launched models—Phi 4 mini reasoning, Phi 4 reasoning, and Phi 4 reasoning plus—are designed as “reasoning” models, allowing for more thorough verification of solutions to complex challenges. This expanded family aims to support developers in creating applications that function at the edge of AI technology.
The Phi 4 mini reasoning model is relatively compact, comprising approximately 3.8 billion parameters and trained on 1 million synthetic math problems generated by the AI startup DeepSeek. With its focus on educational applications, this model is ideally suited for lightweight devices and embedded tutoring systems.
Parameters in AI models are indicative of their problem-solving abilities, and generally, a higher parameter count correlates with better performance. The Phi 4 reasoning model features 14 billion parameters and utilises a combination of high-quality web data and curated examples derived from OpenAI’s o3-mini. This model is particularly suitable for disciplines such as mathematics, science, and coding.
Conversely, Phi 4 reasoning plus is an adaptation of Microsoft’s previous Phi-4 model, tailored to enhance accuracy in specific tasks. Microsoft asserts that this model approaches the performance of the much larger R1 model, which boasts 671 billion parameters. Internal evaluations have revealed that Phi 4 reasoning plus holds its ground against o3-mini in the OmniMath math proficiency test.
All three models—Phi 4 mini reasoning, Phi 4 reasoning, and Phi 4 reasoning plus—are now accessible through the AI development platform Hugging Face, complete with comprehensive technical documentation.
In their announcement, Microsoft highlighted that these new models employ techniques such as distillation and reinforcement learning, alongside high-quality data. This methodology strikes a balance between size and performance, making the models efficient enough to operate in low-latency environments while still maintaining strong reasoning capabilities akin to larger models. Consequently, even devices with limited resources can effectively manage complex reasoning tasks.
Overall, Microsoft’s launch of these new AI models reinforces its commitment to advancing accessible AI technology, facilitating innovation for developers while ensuring robust performance across a range of applications.
Fanpage: TechArena.au
Watch more about AI – Artificial Intelligence
