Tailor a small language model to your organization’s specific needs with the extensive customization Phi models offer.
Get accelerated response times for real-time guidance, autonomous systems, apps requiring low latency, and other critical scenarios using Phi models trained on high-quality data and built for agility.
Integrate Phi models into your existing infrastructure for flexible deployment and seamless management in the cloud, at the edge, and directly on device.
Phi models are developed in accordance with Microsoft AI principles: accountability, transparency, fairness, reliability and safety, privacy and security, and inclusiveness.
FEATURES
Use Phi for generative AI applications
Flexible deployment
Deploy locally for more control over the model and its environment—with or without connection to the cloud.
Ultra low latency
Designed to deliver ultra-low latency, Phi models offer fast response times making them suitable for real-time applications.
Multimodal capabilities
Deliver enhanced versatility across tasks using multimodal Phi models that integrate text, vision, and audio inputs.
Cost-constrained tasks
Designed to be lightweight and resource efficient, Phi models enable developers to create new tools without compromising on performance.
Customization and precision
Boost performance by fine-tuning the models with domain specific data.
Accurate and relevant answers
Generate more coherent, accurate, and contextually relevant outputs with high quality training data and increased context length.
Security
Embedded security and compliance
34,000
Full-time equivalent engineers dedicated to security initiatives at Microsoft.
15,000?
Partners with specialized security expertise. ?
>100
Compliance certifications, including over 50 specific to global regions and countries.
Explore flexible pricing for the family of Phi models. Each model supports various pricing options to fit your needs. Phi models are available through model as a service (MaaS) with pay-as-you-go billing via inference APIs.
Users can also access Phi models for free for real-time deployment through the Azure AI model catalog or Hugging Face.
Phi-4-mini and Phi-4-multimodal are the newest models in the Phi open model family. These two models join Phi-4 as the next models in the series.?
Phi-4-mini features a 200,000 word vocabulary allowing for increased multilingual support, grouped-query attention, built in function calling, improved instruction following, and shared embedding (compared to Phi-3.5-mini).?
Phi-4-multimodal is the first model of the Phi family that supports text, audio, and vision inputs allowing for natural and context-aware interactions. This model features a 200,000 word vocabulary in more than 20 languages, and specializes in speech recognition, translation, summarization, Q&A, and audio understanding. It also provides comprehensive image analysis capabilities, including OCR, chart and table interpretation, and multi-image/multi-frame comparison.
Phi models have evolved since their inception. Starting with Phi-1 and Phi-2, Microsoft introduced groundbreaking innovation in AI, that went on to drastically change the tech landscape. Phi refers to the family of models as a whole. Within the Phi model family, there are multiple series of models. See below for the full list: ?
Phi-4-multimodal: For tasks that require text, audio, and image inputs
Phi-4-mini: For enhanced natural language instruction prompting and advanced reasoning
Phi-4: For complex reasoning and math problem solving?
Phi-3.5: For enhanced multilingual support and improved performance
Phi-3: For language understanding, reasoning tasks, and coding benchmarks
Phi-2: For language comprehension
Phi-1.5: For reasoning and understanding
Phi-1: For Python coding
Phi: For language understanding and reasoning tasks ?
*Phi also performs well on coding benchmarks.
Yes, Phi models have been through rigorous safety post-training. The Phi models, starting from Phi-3 onwards, are designed for production environments. These models are highly capable and cost-effective, making them suitable for a wide range of applications.
Yes, Phi models can be fine-tuned. Read more about fine-tuning Phi in the Phi Cookbook.
Phi was originally trained and optimized for English. The more recent models have been expanded to support over 20 languages.
The newest models were trained on offline datasets with a cutoff date of June 2024 for publicly available data.
Phi is the Microsoft family of small language models (SLMs). These models were created to give developers tools to implement AI directly on a device without the need for cloud connectivity. Phi models are open source through the MIT License.
Phi-3 was the first set of publicly available SLMs released. With the addition of new model offerings, including Phi-3.5 and Phi-4, the series is now unified under the Phi brand name.
Next steps
Choose the Azure account that’s right for you
Pay as you go or try Azure free for up to 30 days.
Get the Azure mobile app 
