SageMaker AI adds serverless customization for Muse Glimmer, Qwen3.8 and Ministral 3

AWS announced on October 22, 2026, that Amazon SageMaker AI now supports serverless model customization for Muse Glimmer, Qwen3.8 27B, and Ministral 3 in 3B, 8B and 14B sizes. The update adds customization to the deployment support already offered for these models, according to AWS’s launch notice.

Training options and workflow

AWS lists supervised fine-tuning (SFT), direct preference optimization (DPO) and reinforcement fine-tuning (RFT) as supported customization options for the newly added models. It also lists low-rank adaptation (LoRA) and full fine-tuning.

The announcement describes SFT as using labeled examples for domain-specific tasks, DPO as using preference data to shape outputs, and RFT as using reward signals.

AWS says SageMaker AI handles infrastructure provisioning and training orchestration. Developers can start customization jobs through the Models page in SageMaker Studio, the SageMaker Python SDK, or AWS’s Agent skills with a coding assistant.

Availability and pricing

The announcement lists serverless model customization as available in four AWS regions: US East (N. Virginia), US West (Oregon), Asia Pacific (Tokyo), and Europe (Ireland).

AWS describes charging as usage-based, but the notice does not include pricing rates.

What engineers should evaluate

The practical change is a broader model selection within AWS’s managed customization workflow—not evidence that a customized model will perform better for a particular application.

Teams should compare an unchanged model against a customized version on held-out examples from the intended workflow. For preference- or reward-based training, evaluation should also check whether gains on the training objective come at the expense of other requirements.

AWS says customization can help smaller models match frontier-model quality on specific tasks while reducing cost and latency. The announcement provides no benchmarks for these additions. Engineers evaluating that proposition should include a larger-model baseline and compare quality, cost and latency under comparable conditions, rather than assume the benefits follow from customization alone.