Microsoft introduced the AI GENERATIVE MODEL PHI-4 With weights fully open on the embraced facial platform. Since its presentation in December 2025, Phi-4 has drawn attention to its improved performance in mathematical calculations and understanding of multitasking language, while requiring less IT resources than larger models.
Phi-4, with 14 billion parameters, is designed to compete with models like GPT-4O Mini, Gemini 2.0 Flash and Claude 3.5 Haïku.
This small language model (SLM) is optimized for complex mathematical calculations, logical reasoning and effective multitasking. Despite its smaller size, Phi-4 offers high performance, deals with long contexts and is ideal for applications that require precision and efficiency. Another remarkable feature is its MIT license, allowing free use, modification and distribution, even for commercial purposes.
Microsoft has further improved the model using synthetic data and fine adjustment techniques, improving its precision in tasks requiring reasoning. An example of Phi-4 mathematical reasoning capacities is demonstrated in the figure below.
In April 2025, Microsoft introduced Phi-3 Mini, the first in the Phi-3 series of small languages models. It included 3.8 billion parameters and was formed on a smaller data set compared to larger models like GPT-4. This was followed in August by Phi-3.5 models, including Phi-3.5-Vision and Phi-3.5-MOE, which used synthetic data and filtered public data sets, supporting contexts up to 128,000 tokens. This evolution in the development of small models led to the release of Phi-4.
Initially, Phi-4 was available via the Foundry Azure Ai platform. Now Microsoft has released Phi-4 on the Face Platform with open weights under a MIT license. Phi-4 is also available via Ollla.
Phi-4 excels in several key areas. It surpasses more important models in mathematical calculations, including solving complex problems, making it ideal for applications that require high precision calculations. It is also very effective in multitasking and logical reasoning, supports long -context treatment and achieves high performance with limited calculation resources. These features make Phi-4 an excellent choice of integration into scientific and commercial projects requiring both the accuracy and optimization of resources.
By making Phi-4 weights accessible to the public, developers can integrate it freely into their projects and personalize it for specific tasks, considerably expanding its potential applications in various fields.
For more details on technical specifications, we recommend that you review the Full technical report on Arxiv.
