Phi-4-reasoning-vision

Phi-4-reasoning-vision-15B, published March 4, 2026, is the newest general-purpose member of Microsoft's Phi small-language-model family. It pairs the Phi-4-Reasoning backbone with a SigLIP-2 vision encoder in a mid-fusion design, and its signature feature is selective reasoning: the model decides when to emit a chain of thought and when to answer directly, keeping token costs down on easy inputs.

Microsoft published the weights, fine-tuning code, and benchmark logs under the MIT license on Hugging Face, GitHub, and Microsoft Foundry. One caveat from the model card: the context window is only 16,384 tokens — the same short window as base Phi-4, not the 128K of Phi-4-multimodal — so it's best suited to image-plus-reasoning tasks rather than long documents. Note: despite third-party rumors, no 'Phi-5' exists as of October 2026; Microsoft's newer releases have been under the MAI brand.