Microsoft’s Phi 4 model – big thinking in a small size, giving tough competition to large machines.
New Delhi – Microsoft took a big step in the world of artificial intelligence by launching its new “open” AI models on Wednesday. These models are not only fast, but also have the ability to solve complex problems on devices with low resources. The most important thing among these is that Microsoft claims that the performance of some of its models is also competing with OpenAI’s popular o3-mini model.
The three new models launched by Microsoft are: Phi 4 Mini Reasoning, Phi 4 Reasoning, and Phi 4 Reasoning Plus. All these models are “reasoning” based — that is, they understand the problems better and fact-check them before solving them. This means that these models are capable of giving more accurate answers.
The Phi 4 Mini Reasoning model has been specially designed for educational applications. It has been trained with about 1 million synthetic (artificial) mathematical questions, which were prepared by the R1 model of Chinese AI company DeepSeek. The model is about 3.8 billion in size. Microsoft says that the model can work as an “embedded tutor” on small and lightweight devices.
The Phi 4 Reasoning model, on the other hand, is much more powerful. It has 14 billion parameters, making it excellent in areas such as math, science and coding. It is trained on high-quality web data and examples from OpenAI’s o3-mini model.
The most advanced version is Phi 4 Reasoning Plus, which is a new version of Microsoft’s existing Phi-4 model. Its reasoning ability has been enhanced so that it can give even more accurate results in specific tasks. Microsoft says that the performance of this model has now come close to DeepSeek’s massive 671 billion parameter R1 model. In addition, according to the company’s internal tests, this model competes with OpenAI’s o3-mini in a mathematical test called OmniMath.
All these new AI models have been released by Microsoft on the popular AI platform called Hugging Face, where developers can use them. The company says that these models have been developed with the help of distillation, reinforcement learning and high-quality data, which keeps their size small but performance strong. This balance makes them capable of working fast even on devices with low latency and limited resources.
Related this : Relevance AI Raises $24M to Accelerate AI Agent Adoption in the Workplace
