Phi-4-multimodal

Claims and evidence

• Released 26 February 2025 ● Primary source[1]page date and first paragraph [2]Training, Model: Release date February 2025

• Status AvailableWeights remain downloadable from Microsoft's Hugging Face organisation; Microsoft Foundry lists the model as GA without a retirement date (checked 2026-10-01). ● Primary source[2]weights and licence (MIT) [4]Foundry Models from partners and community, Microsoft: Phi-4-multimodal-instruct, GA, no retirement date

• Derived from (adapter) Phi-4-miniUses the pretrained Phi-4-mini-instruct language model as backbone, kept frozen, with added vision and speech encoders, projectors and modality-specific LoRA adapters (mixture of LoRAs). ● Primary source[2]Training, Model: Architecture [3]Abstract; vision and speech/audio modality sections
• Successor of Phi-3.5-vision EditorialEditorial link along the Phi multimodal line: the model card presents the release as a response to Phi-3 series feedback, where users had to chain a speech recognition model with the Mini and Vision models, but names no direct predecessor. Phi-3.5-vision was the latest Phi vision model before it. ● Primary source[2]Release Notes

• Change · Modality Adds image and audio input to the Phi-4-mini language model, which stays frozen, through vision and speech encoders with LoRA adapters.Compared with Phi-4-mini ● Primary source[2]Training, Model: Architecture [3]
• Change · Modality Adds speech and audio input; with the Phi-3 series, users had to put a separate speech recognition model in front of the Vision model.Compared with Phi-3.5-vision ● Primary source[2]Release Notes
• Change · Architecture Processes text, vision and speech in a single model with a mixture of LoRAs, without separate pipelines per modality.Compared with Phi-3.5-vision ● Primary source[1]
• Change · Languages Uses a larger vocabulary for multilingual support, with text input in 23 languages; vision input is English only.Compared with Phi-3.5-vision ● Primary source[2]Model Summary

• Input text, image, audio ● Primary source[2]Model Summary; Training, Model: Inputs [1]Natively built for multimodal experiences
• Output text ● Primary source[2]Model Summary; Training, Model: Outputs
• Feature Function calling ● Primary source[2]Primary Use Cases: Function and tool calling
• Feature MultilingualText in 23 languages, vision in English, audio in 8 languages according to the model card. ● Primary source[2]Model Summary: supported languages per modality
• Open weights Yes ● Primary source[2]licence: MIT
• Context window 128K tokens ● Primary source[2]Model Summary; Training, Model: Context length
• Parameters 5.6B ● Primary source[1]What is Phi-4-multimodal? [2]Training, Model: Architecture
• Access API, Open-weights download, cloud partnerAt launch: Azure AI Foundry (api), Hugging Face (open-weights-download) and the NVIDIA API Catalog (cloud-partner). ● Primary source[1]first paragraph; Learn more about Phi-4

Lineage

Variants

No variants recorded in this record.

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from Microsoft on AI Radar →