Microsoft adds 4 open models and shared hosting to Foundry

Microsoft added 4 open-source models to Foundry Managed Fine-Tuning, its tool to train a model on your examples. It also added MAI-Code-1.1-Flash. All of this is in public preview, a test version that is open to all. Microsoft also started Multi-LoRA Managed Compute. There, adapters (small add-on parts) share the weights and GPU capacity of one parent model.

What is new

The 4 models are Muse-Glimmer-30B, qwen3.8-27b, qwen3.6-35b-a3b and gpt-oss-120b. Microsoft names the task for each one:

• Muse-Glimmer-30B: tool calls, also after a result that is not planned.
• qwen3.8-27b: long research and planning tasks.
• qwen3.6-35b-a3b: code work in a repository and in frontend code.
• gpt-oss-120b: hard analysis, with more capacity than gpt-oss-20b.

All 4 models accept supervised fine-tuning, which trains a model on example answers. They also accept reinforcement fine-tuning, which trains a model with a reward.

The proof

Microsoft gives no test results, no prices and no speed numbers. Only Microsoft names the task for each model.

What is not known

Microsoft gives no price for shared hosting. It does not tell how many adapters can use one parent model.

Insight: Microsoft sells the tuning layer around open models

Meta made an open-weight model for use on the hardware of a company. Nvidia showed open models for document search. Microsoft adds open models to Foundry, with shared hosting. This seems a move to own the position where companies tune open models, from any maker. Leaders: open weights do not free you from a cloud vendor.