Boost Your AI Capabilities with Azure’s New Phi-3.5 Model Family

Microsoft is pushing the boundaries of AI innovation beyond its well-known partnership with OpenAI. The company has made a significant leap with the introduction of three new models in its Phi series, including the Phi-3.5-mini-instruct, Phi-3.5-MoE-instruct, and Phi-3.5-vision-instruct, which are designed to enhance reasoning, vision, and multilingual capabilities, all while being accessible under an open-source MIT License. With these powerful models, Microsoft aims to push the boundaries of AI performance, offering developers cutting-edge tools to innovate and build advanced AI applications.

The New Phi-3.5-MoE: Mixture of Experts Model

One of the standout additions to the Phi family is the Phi-3.5-MoE (Mixture of Experts) model. This innovative model architecture is designed to enhance performance and reduce latency while maintaining high-quality outputs. Here's how it works:

• Architecture: The Phi-3.5-MoE model consists of 16 experts, each with 3.8 billion parameters. However, unlike traditional models where all parameters are active at once, the MoE architecture activates only 6.6 billion parameters at a time. This selective activation allows the model to deliver the computational efficiency of a smaller model while retaining the domain expertise and output quality typically associated with much larger models.

• Performance: Despite its selective parameter usage, the Phi-3.5-MoE model boasts a total size of 42 billion parameters. This architecture enables the model to outperform larger, dense models in terms of speed and quality, particularly in multi-lingual tasks. Supporting over 20 languages, the Phi-3.5-MoE model excels in various benchmarks, proving highly competitive even against models with significantly more active parameters.

• Applications: Companies like CallMiner have already begun leveraging the Phi-3.5-MoE model for its speed, accuracy, and security. The ability to fine-tune the model for specific use cases, combined with its robust safety measures, makes it an ideal choice for enterprises seeking to deploy high-performance AI solutions across multiple languages.

Introducing the Phi-3.5-mini: Compact Yet Powerful

Alongside the Phi-3.5-MoE, Azure has also introduced the Phi-3.5-mini model, designed to offer high performance in a compact form factor. Despite its small size, this model packs a punch:

• Model Specifications: The Phi-3.5-mini model is a 3.8 billion parameter model that has undergone extensive pre-training and post-training, using multi-lingual synthetic data. This process has significantly enhanced its multi-lingual capabilities, enabling the model to support over 20 languages with remarkable proficiency.

• Multi-Lingual and Long-Context Capabilities: Phi-3.5-mini excels in tasks that require multi-lingual support and long-context processing. With a context length of up to 128K, it is particularly effective in summarizing long documents, answering questions based on extensive text, and retrieving information from large datasets. This makes it a strong contender in comparison to larger models like the Llama-3.1-8B and Mistral-7B families.

• Use Cases: The Phi-3.5-mini model is ideal for applications requiring multi-lingual support without sacrificing performance. It is especially well-suited for tasks in languages like Arabic, Dutch, Finnish, Polish, Thai, and Ukrainian, where it has shown substantial performance improvements over previous versions.

Enhancing AI Outputs with Guidance

In addition to these models, Azure AI has introduced Guidance, a powerful tool that enhances the predictability and structure of model outputs. Guidance allows developers to define precise programmatic constraints for the model, ensuring that outputs meet specific criteria:

• Structured Output: With Guidance, developers can constrain the model to produce outputs in formats like JSON, Python, HTML, or SQL, making it easier to integrate AI into various applications. This tool can also limit outputs to direct quotes from provided contexts or steer the model to select from predefined lists, significantly improving the quality and relevance of the results.

• Efficiency Gains: By eliminating the need for expensive retries and reducing costs and latency by up to 50%, Guidance makes the Phi-3.5-mini serverless endpoint even more powerful for structured scenarios.

Phi-3.5-Vision: Multi-Frame Image Understanding

Another exciting development in the Phi-3.5 family is the Phi-3.5-vision model, which introduces multi-frame image understanding and reasoning capabilities:

• Multi-Frame Support: This model can analyze multiple input images, allowing for detailed image comparison, summarization, and even video summarization. This capability unlocks new possibilities for applications in fields like security, media analysis, and medical imaging.

• Performance: With 4.2 billion parameters, the Phi-3.5-vision model is optimized for complex visual tasks, offering improved performance over single-image benchmarks.

A Commitment to Safe and Responsible AI

At the core of Azure AI's strategy is a commitment to developing safe and responsible AI. The Phi-3.5 models are no exception, offering a suite of tools to ensure quality and safety:

• Built-In Safety Measures: Developers can assess the quality and safety of their models using Azure AI's evaluation tools, which include built-in controls and custom metrics. Features like prompt shields and protected material detection help safeguard against potential misuse.

• Real-Time Monitoring: Once deployed, developers can monitor their applications in real time for quality and safety issues, including adversarial prompt attacks and data integrity concerns, enabling timely interventions.

Conclusion

The Phi-3.5 model family represents a significant advancement in AI technology, offering a range of models that cater to different needs, from high-performance multi-lingual tasks to efficient and secure applications. With the introduction of the Phi-3.5-MoE and Phi-3.5-mini models, along with tools like Guidance and the enhanced Phi-3.5-vision model, Azure AI continues to lead the way in providing robust, scalable, and responsible AI solutions for enterprises across the globe.

If you want more updates related to AI, subscribe to our Newsletter