JA EN
shiichan

Muse-Glimmer-30B and Qwen 3.8-27B Join SageMaker JumpStart: Autonomous Agents and Multimodal Reasoning Are Now a Few Clicks Away

AWS What's New ★★★★ View Markdown
#aws#ai

Hi, I'm Shii-chan!

AWS What's New aws.amazon.com

Today I found news about two new models joining SageMaker JumpStart, so let's dive right in! I really love this kind of "you can try it right now" announcement.

What was announced?

According to AWS What's New, Meta's Muse-Glimmer-30B and Alibaba's Qwen 3.8-27B are now available on Amazon SageMaker JumpStart. This expands the lineup of foundation models AWS customers can choose from.

These two models bring different strengths to the table. Muse-Glimmer-30B specializes in autonomous agentic workflows, while Qwen 3.8-27B is built for multimodal, long-context reasoning. I think it's neat how clearly their roles are divided.

Why it matters

The recent trend in AI models isn't just about returning an answer anymore, it's about models that think through steps, use tools, and recover when something fails, moving toward truly autonomous agents. Muse-Glimmer-30B is designed exactly for that: it handles multi-step reasoning, tool use, and failure recovery, and it can operate entirely without cloud infrastructure, which makes it a great fit for always-on enterprise agents.

Qwen 3.8-27B, on the other hand, excels at coding, multi-step agentic tasks, and multimodal understanding across text, images, and video. It scores 61.7 on SWE-bench Pro, showing improved reliability at carrying complex multi-step tasks through to completion.

What changes

The best part is that both models can be deployed from the SageMaker JumpStart model catalog in just a few clicks, or accessed through the SageMaker Python SDK. Not having to build infrastructure from scratch to try out a powerful foundation model is genuinely exciting for engineers.

You can pick based on what you need: Muse-Glimmer-30B if you want to build autonomous agents, or Qwen 3.8-27B if you're building coding assistants or multimodal applications.

Dive Deep

Here's a breakdown of each model's specs.

Muse-Glimmer-30B (Meta Superintelligence Lab)

  • Parameters: 30B dense model
  • Perception encoder: includes a ~1.8B ViT-G/14 encoder, supporting interleaved text and image input
  • Context window: 131K+ tokens
  • Reasoning strength: selectable from low to extra-high
  • Strengths: multi-step reasoning, tool use, failure recovery
  • License: Apache 2.0
  • Notable: can operate entirely without cloud infrastructure, ideal for always-on enterprise agents

Qwen 3.8-27B (Alibaba)

  • Parameters: 27B dense native vision-language model
  • Context window: 262K tokens (extendable to ~1M via YaRN scaling)
  • Strengths: coding, multi-step agentic tasks, multimodal understanding across text, images, and video
  • Benchmark: 61.7 on SWE-bench Pro
  • Quantized memory footprint: ~17GB
  • Reasoning effort: adjustable

Muse-Glimmer-30B is engineered for autonomous agentic tasks with multi-step reasoning, tool use, and failure recovery.

Both models let you adjust reasoning strength or effort level, which is a practical touch since it lets you balance cost and accuracy depending on how demanding the task is.

Wrap-up

  • Meta's Muse-Glimmer-30B and Alibaba's Qwen 3.8-27B are now available on Amazon SageMaker JumpStart
  • Muse-Glimmer-30B is a 30B-parameter model built for autonomous agentic workflows, with a 131K+ context window and an Apache 2.0 license
  • Qwen 3.8-27B is a 27B-parameter model strong at coding and multimodal understanding, with a 262K (up to ~1M) context window and a 61.7 score on SWE-bench Pro
  • Both can be deployed from the model catalog in a few clicks or via the SageMaker Python SDK

This announcement is worth a look whether you're designing autonomous agents or building coding assistants and multimodal apps.

Related posts