Microsoft says Foundry Managed Compute is generally available for serving open models on dedicated GPUs. The company provisions the GPUs, maintains the containers and operates the serving software for deployments made through Foundry.
The service launched in public preview in June 2026. Microsoft says its catalog now offers more than 150 open models for Managed Compute, including recent additions for reasoning, coding, search, document processing and speech.
Teams choose a model and a deployment template that sets its context length, serving software and GPU configuration. Depending on the model, they can use NVIDIA A100, H100 or H200 GPUs, or AMD MI300X GPUs. Microsoft has also added deployment options in United States and European Union Data Zones. The deployment wizard shows the hourly cost before launch.
Teams can add instances behind an existing endpoint as traffic grows. Each instance uses the number of GPUs in its template. Microsoft says a DeepSeek V4.1 Flash (1017) deployment using four MI300X GPUs per instance would grow from four to 20 GPUs when scaled from one instance to five. Customers pay hourly for each accelerator.

Microsoft says it reviews model licenses, screens for untrusted code and validates model weights before adding them to the catalog. It also says it builds, scans and signs the serving images, then rolls out software updates to existing deployments.
Custom model serving enters preview
Microsoft also announced preview options for custom models. Teams can deploy registered LoRA adapters produced by Foundry fine-tuning when they match a base-model version and a template that supports LoRA. Multiple adapters can share the GPUs of the deployment hosting them, with a separate deployment name for each.
Bring Your Own Weights is in limited public preview. It lets teams register and deploy custom full-weight SafeTensors models, including models trained outside Foundry, using a compatible catalog base model and a Microsoft-provided template. Microsoft is accepting sign-ups for early access.



