Microsoft’s most capable new Phi 4 AI model rivals the performance of far larger systems

Microsoft launched a number of new “open” AI fashions on Wednesday, essentially the most able to which is aggressive with OpenAI’s o3-mini on no less than one benchmark.

All the new pemissively licensed fashions — Phi 4 mini reasoning, Phi 4 reasoning, and Phi 4 reasoning plus — are “reasoning” fashions, which means they’re capable of spend extra time fact-checking options to advanced issues. They increase Microsoft’s Phi “small mannequin” household, which the corporate launched a yr in the past to supply a basis for AI builders constructing apps on the edge.

Phi 4 mini reasoning was educated on roughly 1 million artificial math issues generated by Chinese language AI startup DeepSeek’s R1 reasoning mannequin. Round 3.8 billion parameters in measurement, Phi 4 mini reasoning is designed for academic purposes, Microsoft says, like “embedded tutoring” on light-weight units.

Parameters roughly correspond to a mannequin’s problem-solving abilities, and fashions with extra parameters typically carry out higher than these with fewer parameters.

Phi 4 reasoning, a 14-billion-parameter mannequin, was educated utilizing “high-quality” internet information in addition to “curated demonstrations” from OpenAI’s aforementioned o3-mini. It’s greatest for math, science, and coding purposes, in response to Microsoft.

As for Phi 4 reasoning plus, it’s Microsoft’s previously-released Phi-4 mannequin tailored right into a reasoning mannequin to attain higher accuracy on explicit duties. Microsoft claims that Phi 4 reasoning plus approaches the efficiency ranges of R1, a mannequin with considerably extra parameters (671 billion). The corporate’s inside benchmarking additionally has Phi 4 reasoning plus matching o3-mini on OmniMath, a math abilities take a look at.

Phi 4 mini reasoning, Phi 4 reasoning, and Phi 4 reasoning plus can be found on the AI dev platform Hugging Face accompanied by detailed technical studies.

“Utilizing distillation, reinforcement studying, and high-quality information, these [new] fashions stability measurement and efficiency,” wrote Microsoft in a weblog submit. “They’re sufficiently small for low-latency environments but keep robust reasoning capabilities that rival a lot larger fashions. This mix permits even resource-limited units to carry out advanced reasoning duties effectively.”