Microsoft presents Phi-4: new compact models with high reasoning skills | Best large language models in the world | Large language models tutorial python github | Large language models | Turtles AI

Microsoft has recently expanded its range of AI models with the launch of the Phi-4 series, designed to offer advanced reasoning capacity in compact formats. These models, available on Hugging Face with permissive license, include Phi-4 Mini Reasoning, Phi-4 Reasoning and Phi-4 Reasoning Plus, each optimized for specific application areas.

Key points:

• Compact models, high performance: Phi-4 Mini Reasoning, with 3.8 billion parameters, exceeds models of similar sizes and approaches the performance of larger models in math and programming tasks.
• Optimization for reasoning: Phi-4 Reasoning, with 14 billion parameters, was trained on high quality web data and well-kept demonstrations, offering competitive performance in mathematics, sciences and programming.
• MULIMODAL approach: Phi-4 Multimodal integrates text, vision and audio, using Lora adapters to manage different input modes without interference.
• Competitive benchmarking: Phi-4 Reasoning Plus approaches the performance of significantly larger models, such as Deepseek R1, in benchmark as Omnimath.

Phi-4 Mini Reasoning was trained on about one million synthetic mathematical problems generated by the R1 reasoning model of the Chinese Deepseek startup. Despite its small size, this model is designed for educational applications, such as integrated tutoring on light devices.

Phi-4 Reasoning, with 14 billion parameters, was trained using high quality web data and curated demonstrations, making it ideal for math applications, sciences and programming. Phi-4 Reasoning Plus, an improved variant, has been optimized through a learning phase for reinforcement based on the results, offering higher performance in specific tasks.

Phi-4 MultiDal extends the skills of previous models by integrating text, vision and audio in a single model. Using Lora adapters and specific routers in mode, this model supports scenarios involving visual, linguistic and audio input combinations, overcoming vision-cail and largest vision and language models in a wide range of tasks.

The Phi-4 models have been evaluated through benchmark in mathematics, scientific reasoning, programming, resolution of algorithmic problems, planning and space understanding. In particular, Phi-4 Reasoning Plus approaches the performance of Deepseek R1, which has a significantly greater number of parameters (671 billion), in benchmark as Omnimath.

These models represent a balance between compact dimensions and high performance, allowing the efficient execution of complex reasoning activities also on devices with limited resources.

With the introduction of the Phi-4 series, Microsoft continues to push the boundaries of the abilities of the AI ​​models, offering versatile and powerful solutions for a wide range of applications.