Phi-4-Mini and Phi-4-Multimodal
3.8B Phi-4-Mini with 200K vocabulary, and Phi-4-Multimodal handling text, vision and speech through Mixture-of-LoRAs.
Modality-specific LoRA adapters and routers let one backbone take vision, audio and text without interference; its 460M-parameter audio adapter ranked first on the OpenASR leaderboard per the report.
- Date
- Monday, 3 March 2025
- Lab
- Microsoft
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| OpenASR leaderboard | rank 1 460M-parameter audio LoRA; per technical report | company |
Dated by arXiv v1 (2025-03-03); Microsoft's launch blog and Hugging Face release came a few days earlier (late Feb 2025), not opened here.
Sources
This record was checked against its sources on 6 October 2026. How we check