Microsoft Open-Sources Phi-4 AI Model with Full Weights

The 14-billion-parameter model, now available on Hugging Face under an MIT License, outperforms larger competitors in key benchmarks.

Microsoft has fully open-sourced its Phi-4 large language model, including its weights, making it available for public and commercial use on Hugging Face.
Phi-4, trained on 14 billion parameters, excels in mathematical reasoning, multitask language understanding, and functional code generation, outperforming larger models like OpenAI's GPT-4o-mini and Meta's Llama-3.3.
The model was trained on 9.8 trillion tokens, combining high-quality public data, synthetic datasets, and academic resources, with a focus on math, coding, and logic.
Phi-4's efficient architecture reduces computational and energy demands, providing advanced AI capabilities to developers and organizations with limited resources.
Microsoft emphasizes the importance of safety and alignment, incorporating measures to address bias and misinformation, while encouraging developers to implement additional safeguards for sensitive applications.