Microsoft previews hosting customers’ own model weights in Foundry

Microsoft announced a limited public preview on November 20 that lets eligible Foundry customers deploy their own full model weights on its Managed Compute service. The option accepts models in the SafeTensors format, including weights trained outside Foundry. Deployments require a compatible model from the Foundry catalog and a Microsoft-provided template.
Microsoft had listed custom weights on its roadmap in June. The new preview is separate from general availability of Managed Compute itself, which Microsoft announced at Ignite on November 17. That service hosts open models on dedicated GPUs, while Microsoft operates the serving software and maintains the containers. Customers pay by the GPU-hour. For catalog models, templates specify the context length, serving software and GPU configuration.
The preview gives teams that already trained or fine-tuned compatible models elsewhere another way to put those weights into service without maintaining the underlying GPU infrastructure. It does not make custom model serving generally available or remove the requirement to use a compatible base model.
Sources and citations
Revision history
Published .