JA EN
shiichan

Qwen3's search and reranking models land in SageMaker JumpStart!

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

Hi everyone, it's Shiichan! I found some retrieval-related model news today, so let's dive right in!

AWS What's New aws.amazon.com

What was announced?

According to AWS's What's New, Qwen3-VL-Embedding-2B and Qwen3-Reranker-4B, both built by the Qwen team, are now available in Amazon SageMaker JumpStart. These two models are specialized for information retrieval, giving you a new option if you want to build a search pipeline on AWS.

Why it matters

These two models play different roles in a retrieval pipeline.

  • Qwen3-VL-Embedding-2B handles the "initial recall" stage, quickly narrowing down candidates
  • Qwen3-Reranker-4B handles the "reranking" stage, refining those candidates further

This two-stage setup is a common pattern in real-world search systems, and pairing Qwen3-VL-Embedding-2B with Qwen3-Reranker-4B makes it easier to build a retrieval pipeline that balances accuracy and speed. The embedding model also isn't limited to text: it accepts images, screenshots, videos, and mixed-modality inputs, so it opens the door to use cases beyond plain text search.

What changes

Building these kinds of models yourself used to mean picking a model and setting up hosting infrastructure, which took real effort. Now that they're in SageMaker JumpStart, you can deploy them with a few clicks from the Models section of SageMaker Studio, or programmatically via the SageMaker Python SDK. Being able to bring in multimodal search and reranking without building infrastructure from scratch is a nice win.

Dive Deep

Let's look at what each model supports.

Qwen3-VL-Embedding-2B handles these inputs and tasks:

  • Accepts text, images, screenshots, videos, and mixed-modality combinations of these
  • Generates embeddings that place visual and textual information into a shared vector space
  • Supports image-text retrieval, video-text matching, visual question answering, and multimodal content clustering
  • Supports over 30 languages

Qwen3-Reranker-4B handles these roles:

  • Takes a query-document pair and outputs a relevance score to refine retrieval results
  • Supports text retrieval, code retrieval, text classification, text clustering, and bitext mining
  • Supports over 100 languages, with user-defined instructions to tune performance for specific tasks, languages, or scenarios

Deployment is straightforward: you can deploy to your own AWS account by clicking through the Models section of SageMaker Studio, or with code via the SageMaker Python SDK. For detailed steps, check the Amazon SageMaker JumpStart documentation. The announcement doesn't mention pricing or region availability, so check the documentation for those details before you try it.

Wrap-up

  • Qwen3-VL-Embedding-2B and Qwen3-Reranker-4B are now available in Amazon SageMaker JumpStart
  • The embedding model handles multimodal initial recall, while the reranker handles precise reranking, forming a two-stage pipeline
  • Qwen3-VL-Embedding-2B supports over 30 languages, and Qwen3-Reranker-4B supports over 100 languages
  • You can deploy either one easily from the Models section of SageMaker Studio or the Python SDK

This is great news for engineers who want to build search systems or multimodal RAG pipelines on AWS!

Related posts