Microsoft launches the Phi-4-multimodal and Phi-4-mini models to provide advanced the developers

Microsoft  has announced two new models of small language (SLM) of the Phi family, Phi-4-multimodal and Phi-4-mini, which are designed to provide developers with artificial intelligence (AI) abilities advanced in voice processing tasks, text and images.

The company has commented on its blog that these models are “Ideal for extreme scenarios and environments with computing limitations“And they are already being used to explore different cases of use in various industries, including the detection of anomalies in the manufacture or improvement of customer experiences in retail trade.

Phi-4-multimodal It is a model of 5,600 million parametersthat integrates voice, images and text processing in a single unified architecture. Because it takes advantage of advanced intermodal learning techniques, it allows “more natural and sensitive to context interactions, as Microsoft explained.

This model can process both images and audio at the same time, as well as graphics, tables and documents, surpassing specialized models such as whisperv3 and SeamlessM4T-V2-Large, in automatic voice and translation recognition tasks. In addition, it offers an inference of low latency and high efficiency while the execution is optimized in the device and the computational overload is reduced.

Microsoft has also pointed out that Phi-4-multimodal “is among the few open models that successfully implement the voice summary” and reach GPT-4O performance levels. However, he presents A gap with other models such as Gemini-2.0-Flash In tasks for questions to questions, so you are working to improve this capacity in the next iterations.

Ph-4-minfor its part, it is a model of 3.8 billion parameterswith an ability to respond to grouped consultations, a vocabulary of 200,000 words and “is designed for speed and efficiency.” In this sense, the march has indicated that, although it has a compact size, it continues to surpass the largest models in text based on text, included Reasoning, mathematics, instructions monitoring and coding.

This model admits sequences of Up to 128,000 tokens and provides high precision and scalability so that developers can adapt it in advanced AI applications. In addition, it can be integrated into structured programming interfaces.

Microsoft final Azure AI Foundry, Hugging Face y Nvidia API Catalogwith a complete multimodal experience.