Programmed Training: You write the loop, Foundry runs the GPUs

Today, we're introducing Programmed Training in Microsoft Foundry, now in public preview. Programmed Training provides developers with a Python API for authoring custom training recipes: your code decides what to sample, how to score responses, and when to update model weights, while Foundry executes the model operations on serverless GPUs. The preview includes the Fine-Tuning Sessions Python SDK , a public cookbook of training recipes , and LoRA training for four open-weight models: Qwen3.8-27B, Qwen3.6-35B-A3B, gpt-oss-120b, and Muse-Glimmer-30B . Programmed Training joins recent Foundry fin