SageMaker now lets you fine-tune Qwen3.8 and Ministral 3 without a cluster

Amazon has extended serverless model customization in SageMaker AI to Muse Glimmer, Qwen3.8 27B and Ministral 3 in 3B, 8B and 14B sizes, according to an official announcement dated October 22. AWS provisions and orchestrates the training, so customers do not manage a cluster.

Three tuning methods, two ways to train

Customers can adapt the models with supervised fine-tuning (SFT), direct preference optimization (DPO) or reinforcement fine-tuning (RFT). Each can run as low-rank adaptation (LoRA) or as full fine-tuning.

AWS describes the use cases this way: labeled data with SFT for domain accuracy, preference data with DPO to match an organization’s tone, and reward signals with RFT for tasks the model has not handled before. Billing covers only the resources a job actually consumes.

A job can be started from the Models page in SageMaker Studio, through the SageMaker Python SDK, or with AWS’s agent skills inside a coding assistant. Four AWS regions host the feature: Northern Virginia and Oregon in the United States, Tokyo in Asia Pacific, and Ireland in Europe. Nova, Gemma 4, Nemotron 3.5, Qwen, Llama, gpt-oss and DeepSeek models were already supported, and the new ones are added alongside them.

AWS’s pitch is that a smaller, right-sized model trained on proprietary data can match frontier-model quality on a customer’s own tasks, with lower cost and latency.

Longer training inputs arrived a day earlier

The launch follows a separate AWS post from October 21. It says serverless jobs previously used a fixed sequence length per recipe, and that customers can now set it themselves. For SFT and DPO the setting is dataset_max_len; for RFT it is max_prompt_length and max_response_length. AWS says maximums reach up to 16K tokens for many models, depending on model and technique, and sends readers to its documentation for per-model values.

That post lists gpt-oss, Gemma, Llama, Nemotron and Qwen as supported. It does not name Muse Glimmer or Ministral 3, so whether the new models get the same limits is not stated.

What the announcement leaves unanswered

The page names the models but does not say who makes Muse Glimmer, how large it is, or under what license it is offered. Our roundup of recent model releases covers Meta’s Muse Wonder and Muse Spark Open but does not mention Glimmer, so any link to Meta’s Muse line is unconfirmed here. Teams should check the model’s terms before training on it.

No prices appear on the page. The pay-as-you-go wording does not say whether billing is per token, per hour or per job, and nothing is said about the cost of full fine-tuning against LoRA. Full tuning of a 27B model is a heavier job than a LoRA run on a 3B one, so the bill could differ widely.

The quality claim is AWS’s own. No benchmark compares a tuned small model with a frontier model, and the announcement offers no customer result. The effect of tuning depends on data quality. In an AWS experiment we reported earlier, SFT lifted gpt-oss-20B from reasoning in the requested language 22% of the time to 100% on a 50-case test, yet answers were in the right language only 62% of the time. Even a clear gain can leave gaps.

Region coverage matters for some buyers. Ireland is the only European location on the list, so teams that need data kept in other European countries have no option yet. AWS does not say if more regions are coming.

For developers already using SageMaker, the practical change is a wider menu of open-weight starting points, with Qwen3.8 27B as the largest named in this release and Ministral 3 offering three sizes for cost and latency tradeoffs. Whether the serverless route is cheaper than renting GPUs directly is something the announcement does not settle.