Microsoft launches new 'reasoning' AI models to rival OpenAI, Google
What's the story
Microsoft has unveiled a new series of "open" artificial intelligence (AI) models, the most advanced of which is said to be competitive with OpenAI's o3-mini on at least one benchmark.
The newly launched models include 'Phi 4 mini reasoning,' 'Phi 4 reasoning,' and 'Phi 4 reasoning plus.'
All of these "reasoning" models can spend more time fact-checking solutions to complex problems.
The new models have been made available on the AI dev platform Hugging Face, along with detailed technical reports.
Features
Phi 4 mini reasoning: A compact educational tool
The Phi 4 mini reasoning model is intended for educational use cases like "embedded tutoring" on lightweight devices.
It was trained on some one million synthetic math problems created by DeepSeek's R1 reasoning model.
The model features around 3.8 billion parameters, which roughly translate to its problem-solving capabilities, with more parameters meaning better performance.
Information
Phi 4 reasoning: Ideal for math, science, and coding
The Phi 4 reasoning model, which comes with 14 billion parameters, was trained on "high-quality" web data and "curated demonstrations" from OpenAI's o3-mini. Microsoft says this model is ideal for math, science, and coding applications.
Performance
Phi 4 reasoning plus: R1-like high-performance AI model
The Phi 4 reasoning plus is an adaptation of Microsoft's earlier released Phi-4 system into a reasoning model to deliver better accuracy on certain tasks.
Microsoft says that this model comes close to the performance levels of R1, which has a lot more parameters (671 billion).
The company's internal benchmarking has also matched Phi 4 reasoning plus with o3-mini on OmniMath - a math skills test.