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

Importance Score: 65 / 100 🔴

This week, Microsoft revealed multiple novel “open” artificial intelligence models, with the most formidable rivaling OpenAI’s O3-Mini on a specific benchmark.

Introduction to Microsoft’s Innovative AI Models via Permissive Licensing

Each of these newly introduced, unrestricted models—Phi 4 Mini Reasoning, Phi 4 Reasoning, and Phi 4 Reasoning Plus—falls under the “reasoning” model category. These models are engineered to allocate more time to verifying solutions to intricate challenges. These additions broaden Microsoft’s Phi “small model” lineup, initially launched a year prior to assist AI developers in constructing edge applications.

Key Features of Phi 4 Mini Reasoning

Phi 4 Mini Reasoning underwent training using approximately one million synthetic mathematical problems, generated by R1, a reasoning model developed by DeepSeek, a Chinese AI startup. With around 3.8 billion parameters, Phi 4 Mini Reasoning is tailor-made for educational purposes, such as incorporated tutoring on lightweight devices.

Understanding Parameters in AI Models

In general, parameters correlate with a model’s problem-solving capabilities, with higher parameter counts typically resulting in superior performance.

Applications of Phi 4 Reasoning

Phi 4 Reasoning, boasting 14 billion parameters, was trained on superior web data and refined examples from OpenAI’s O3-Mini. Microsoft recommends this model for applications in mathematics, science, and coding.

Enhancements in Phi 4 Reasoning Plus

Microsoft’s Phi 4 Reasoning Plus is an adaptation of its previously released Phi-4 model, transformed into a reasoning model to enhance accuracy on specific tasks. According to Microsoft, Phi 4 Reasoning Plus approaches the efficiency of R1, a model with a significantly greater parameter count (671 billion). Internal benchmarking also shows Phi 4 Reasoning Plus aligning with O3-Mini on the OmniMath assessment, a test designed to gauge mathematical proficiency. Otherwise known as reasoning skills.

Accessibility and Support for AI Developers

Accessible on Hugging Face, a platform for AI developers, Phi 4 Mini Reasoning, Phi 4 Reasoning, and Phi 4 Reasoning Plus come with comprehensive technical reports to aid developers in deployment.

“By leveraging distillation, reinforcement learning, and high-quality data, these [new] models harmonize size and performance,” Microsoft stated in their blog. “They are compact enough for low-latency settings while maintaining robust reasoning abilities that compete with much larger models. This equilibrium enables even resource-constrained devices to accomplish intricate reasoning tasks effectively.”


🕐 Top News in the Last Hour By Importance Score

# Title 📊 i-Score
1 US 'obliterates' Iran's nuclear sites as defiant Tehran strikes back at Israel: Inside the crucial hours when Trump went 'dark' and unleashed hell 🟢 85 / 100
2 From shipping, to proxies, to targeting US bases, Iran’s options to strike back are limited 🔴 75 / 100
3 When prompted to show a female investor, here's what AI created! 🔴 72 / 100
4 So many young women who've never smoked are getting lung cancer. Now scientists think they've found two shocking causes – and reveal what to look out for 🔴 72 / 100
5 9 Foods for Relieving Headaches and Migraines, According to Doctors and Dietitians 🔴 72 / 100
6 John Brown's belly rumbling is small beer, writes Gary Keown, it's officials who can't tell when the ball's over the line that we really need to talk about 🔴 65 / 100
7 When Dubai almost became a part of India 🔴 65 / 100
8 South Korea to raise concerns to US over potential curbs on chipmakers’ China operations — TradingView News 🔴 65 / 100
9 Character.AI taps Meta’s former VP of business products as CEO 🔴 65 / 100
10 Dear Abby: My family thinks I’m making up my long-distance boyfriend 🔵 40 / 100

View More Top News ➡️