Microsoft adds Phi-4 multimodal and Phi-4 mini as it deepens its push into smaller AI models

Microsoft said on Feb. 26, 2025, that it was adding Phi-4-multimodal and Phi-4-mini to its Phi family of small language models, extending a product line it uses to promote efficient AI systems for developers and enterprise customers. The company framed the release as an answer to demand for models that can do useful work without the compute footprint of the largest chat systems.

Phi-4-multimodal is Microsoft’s first multimodal Phi model, meaning it is designed to handle speech, vision and text in a single architecture. Microsoft said the model has 5.6 billion parameters and is intended for low-latency inference and on-device or edge deployment, where speed and resource use matter as much as raw scale.

Phi-4-mini is the smaller text-focused companion in the release. Microsoft described it as a 3.8 billion-parameter dense transformer with grouped-query attention, a 200,000-token vocabulary and a 128,000-token context window, features that are meant to make compact models more capable on long documents, reasoning tasks and function calling.

The company said both models are available through Azure AI Foundry, Hugging Face and the NVIDIA API Catalog. That distribution matters because it puts the models in front of developers already building on Microsoft’s cloud and partner ecosystems, rather than keeping them confined to a research demo.

Microsoft also used the announcement to stress performance claims. It said Phi-4-multimodal reached the top spot on the Hugging Face OpenASR leaderboard with a 6.14% word error rate, and that Phi-4-mini continues to outperform larger models on several text benchmarks. Those claims position Phi as part of Microsoft’s broader case that smaller models can still be commercially useful.

The post focused on capabilities, access points and benchmark results, not on customer adoption figures, revenue impact or a wider rollout schedule. In practical terms, the release indicates that Microsoft wants Phi to sit alongside its larger AI offerings as a cheaper, more flexible option for speech, vision and text workloads.