Microsoft Foundry will now integrate Hugging Face’s gated models, giving enterprises secure steps access to advanced open-source AI models directly within their Azure environment. New gated models will be rolling out this month in Microsoft Foundry, providing customers with the new innovations from the open-source community.

Why Gated Models Matter

Gated models require users to request and receive approval before access for responsible use. By supporting gated models in Foundry and integrating directly with the Hugging Face user access token, we have simplified the process while allowing customers to stay aligned with each provider’s gating and licensing policies. This gives Foundry users access to a more diverse and up-to-date catalog of high-quality open source models and allow them to deploy the models directly in Foundry.

How Hugging Face Tokens Enable Secure and Governed Access

To access gated models, users authenticate using a Hugging Face user access token. This token is tied to an individual, allowing Foundry to verify that the user requesting the model deployment in Foundry has the required permissions from the model publisher on the Hugging Face Hub. 

For organizations seeking stronger oversight on their user’s token, Hugging Face Team and Enterprise Plans offer enhanced token governance capabilities to companies.

How Gated Model Access Works in Microsoft Foundry

Foundry uses Hugging Face access tokens and secret injection to provide access to gated models. Through secret injection, Hugging Face will use the provided access token to verify whether the user has access to the gated model. Once verified, the model will then be securely downloaded and deployed to the user's online endpoint.

How to gain access:

1. 1. Discover: Users browse the Foundry catalog; gated models appear alongside others.
   2. Request Access: A header indicates the model is gated with a link to the model page on Hugging Face for requesting access.
   3. Provide Token: Users create a connection in Microsoft Foundry or a workspace connection in Azure Machine Learning. The custom connection named HuggingFaceTokenConnection, should be created with the key HF_TOKEN and their read or fine-grained token as the value.
   4. Deploy: Create an endpoint with “enforce_access_to_default_secret_stores” set to “enabled”. Once validated, users can deploy the model with enterprise-grade security to the newly created endpoint.

Note: If you share your Microsoft Foundry project or workspace with others, consider deleting the custom key after deployment.  

What models are coming first?

Segment Anything Model 3 (SAM 3)

Segment Anything Model 3 from Meta introduces "Promptable Concept Segmentation," which allows users to segment objects in images using text prompts. This feature enables segmentation of objects that may not be explicitly labeled, supporting open-vocabulary segmentation tasks.

Compared to previous versions, SAM 3 demonstrates improved accuracy, achieving 75–80% of human performance on the SA-Co benchmark suite. Documented use cases for SAM 3 include medical imaging, robotics, augmented and virtual reality, and automated content moderation, where precise identification and separation of objects in images are required.

Roblox PII Classifier

Roblox PII Classifier is designed to detect personally identifiable information (PII) in chat messages, supporting privacy and safety in online interactions. This model was trained on anonymized Roblox chat data and is open-sourced for broader use. It differs from prior approaches by leveraging a large, diverse dataset and incorporating multilingual support, which enables detection of PII across various languages and contexts.

Benchmarks reported by Roblox indicate the model achieves a 94% F1 score for PII detection on production data. Documented use cases include gaming environments, social platforms, and enterprise collaboration tools, where real-time identification of sensitive information is required to help organizations meet compliance and safety standards.

FLUX.1 Schnell

FLUX.1 Schnell is a text-to-image model developed by Black Forest Labs, designed to generate high-quality images in a small number of inference steps. The model produces images in 1 to 4 steps and achieves top ELO scores for visual fidelity. Flux.1 Schnell optimizes for both speed and image quality, enabling rapid generation without requiring extensive compute resources. The model is ideal for creative workflows, marketing automation, and rapid prototyping, where users need to produce visual assets efficiently and iterate quickly on design concepts.

EuroLLM-9B-Instruct

EuroLLM-9B-Instruct is a language model built to handle the linguistic diversity of Europe, supporting over 30 languages and designed with instruction-following capabilities in mind. The model’s performance on the MMLU-Pro benchmark and machine translation tasks highlights its ability to manage complex language understanding and generation.

For organizations operating across borders, EuroLLM-9B-Instruct can help streamline customer support, automate compliance workflows, and adapt content for local audiences. Its broad language coverage and focus on practical tasks set it apart from earlier models that may have been limited to fewer languages or less robust instruction handling.

Bielik‑11B‑v3.0‑Instruct 

Bielik-11B0v3.0-Instruct is an instruction‑tuned generative language model designed to support high‑quality multilingual text understanding and generation across European languages. Trained on texts spanning 32 languages—with a strong focus on Polish—the model enables precise task‑following behavior for use cases requiring nuanced linguistic reasoning and controlled responses in multiple linguistic contexts.

Built on top of the Bielik11Bv3 base model, this version reflects advancements in data curation and supervised finetuning enabled through largescale high performance computing infrastructure. Documented applications include multilingual content generation, information extraction, semantic analysis, and conversational AI, particularly in scenarios where translation of Polish and other European languages is critical for enterprise and research workloads.

New models will be available through Microsoft Foundry on a rolling basis, if there is another gated model you want to see added to catalog, please let us know!

Getting started

Ready to explore these advanced AI models and accelerate your projects? Visit Microsoft Foundry today to access the models listed above from Hugging Face and many more. 

This blog was written with the help of: 

Alvaro Bartolome <alvaro.bartolome@huggingface.co>

Simon Pagezy <simon.pagezy@huggingface.co>

Jeff Boudier <jeff@huggingface.co>

Juan Julian Cea <juan.moran@huggingface.co>