3.84B (Phi-4-mini, dense); 5.57B (Phi-4-multimodal, Mixture-of-LoRAs)
131,072 tokens
MIT
Same-day release of the compact Phi-4-mini-instruct text model and Phi-4-multimodal-instruct. Both use a 32-layer language backbone with a 3,072-dimensional hidden state, 24 attention heads, eight key-value heads, and a 200,064-token vocabulary. Phi-4-mini is a dense decoder-only Transformer with grouped-query attention and shared input-output embeddings, designed for multilingual reasoning, mathematics, coding, instruction following, and function calling in compute-constrained or latency-sensitive environments. Its model card lists support for 24 text languages and describes post-training with supervised fine-tuning and direct preference optimization. Phi-4-multimodal extends the language backbone through a Mixture-of-LoRAs architecture that unifies text, image, and audio inputs while generating text output. It supports simultaneous audio-and-vision reasoning, image and document understanding, OCR, charts and tables, multiple-image comparison, speech recognition, speech translation, speech question answering, speech summarization, and general audio understanding. Text supports the same 24-language set; vision is documented for English, while audio supports English, Chinese, German, French, Italian, Japanese, Spanish, and Portuguese. Multimodal post-training used supervised fine-tuning, direct preference optimization, and reinforcement learning from human feedback. At launch Microsoft reported a 6.14% word-error rate on the Hugging Face OpenASR leaderboard, performance above WhisperV3 and SeamlessM4T-v2-Large on its reported speech-recognition and translation evaluations, and competitive document, chart, OCR, mathematical, and scientific visual reasoning despite the 5.6B model size. Both models target on-device and edge deployment and are released with standard Transformers checkpoints as well as Microsoft-supported deployment paths.