Today, Microsoft Build will introduce a wide range of new features and collaborations, with Microsoft and Hugging Face deepening strategic collaborations and making open models and open source AI easier to use anywhere. Together we will work to enable AI builders across open science, open source, cloud, hardware and developer experiences.
Cloud AI Builder Collaboration
We look forward to unveiling two major new experiences to build AI with open models in Microsoft Azure.
Expanded the HF collection in Azure Model catalog
A year ago, HugFace and Microsoft unveiled a collection of hugging faces in the Azure Model catalog. The embracing face collection has been used by hundreds of Azure AI customers, and since its introduction, over 1,000 open models have been available. Today, we’re adding some of the most popular and open big language models to Hugging Face Collection to allow for a one-click deployment directly from Azure AI Studio.
The new models include Meta’s llama 3, Mistral AI’s Mistral 7b, Cohere for ai’s Command R Plus, QWEN’s QWEN 1.5 110b, and some of the most performant fine-tuning models on the Hagging Face Community’s Open LLM leaderboard.
To deploy your model to your own Azure account, you can start with the model card in HugFace Hub and select the “Deploy Azure” option.
Alternatively, you can find the model directly in Azure AI Studio within the hugging hugging face collection, and click Deploy.
Build AI with the new AMD MI300X in Azure
Today, Microsoft has created a new Azure ND MI300X virtual machine (VM) that is commonly available in Azure, based on the latest AMD Instinct MI300 GPUs. We worked with AMD and Microsoft to achieve the incredible performance and cost/performance of hugging face models in a new virtual machine.
This work leverages deep collaboration between AMD and Optource Library Optimum-AMD, providing continuous testing that embraces optimization, ROCM integration, and face open source libraries and models on AMD Instinct GPUs.
Open Science Collaboration
Microsoft has released some of the most popular open models with hugging faces, and nearly 300 models are currently available on Hug Face Hub in Microsoft organizations.
This includes models from the recent PHI-3 family that may be permitted under MIT and offer performance beyond the weight class. For example, the PHI-3 Mini surpasses many of the large language models with only 3.8 billion parameters, with a large parameter of 7-100 billion, making the model a great candidate for on-device applications.
To demonstrate the functionality of the PHI-3, hugging your Face Deployed Phi-3 Mini, chat with the largest open model in a free consumer application and create an assistant.
Open Source Collaboration
Embracing Faces and Microsoft have collaborated for three years to make face models easy to export and use at the ONNX runtime through the best open source library.
Recently, Hugging Face and Microsoft have focused on the possibility of local inference via WebGPU, leveraging Transformers.JS and the ONNX runtime Web. Learn more about the collaboration of this community article by the ONNX Runtime team.
To check how your WebGPU works, consider a demo of PHI-3, which locally generates 70 or more tokens in your browser!
The browser does not support video playback.
Developer collaboration
Last but not least, today we are unveiling a new integration that makes it easier than ever to build AI applications with embracing facespaces and VS code.
The Hugging Face Community has created over 500,000 AI demo applications in the hub. The new Spaces Dev mode allows you to easily connect spaces to local VS codes by hugging face users, or spin up hosted and web host code environments.
Spaces Dev Mode is currently in beta and is available for Pro Subscribers. For more information about Spaces Dev Mode, see Introducing Spaces Dev Mode for a seamless developer experience and documentation.
What’s next?
We are excited to deepen our strategic collaboration with Microsoft and make open source AI more accessible anywhere. Stay tuned as we enable more models in the Azure AI Studio model catalog and showcase new features and experiences in the coming months.