Qwen3 embedding and reranking models for retrieval are now available in Amazon SageMaker JumpStart
Build multimodal, multilingual search pipelines on AWS using a vision-language embedder and instruction-tunable reranker, deployable in minutes via JumpStart.
View original announcement →Visual Summary
graph TD
A{{Qwen3 Models on SageMaker JumpStart}}:::announced
B(SageMaker Studio):::compute
C([Qwen3-VL-Embedding-2B]):::feature
D([Qwen3-Reranker-4B]):::feature
E(SageMaker Endpoint):::compute
F((User/Application)):::external
G([Multimodal Inputs]):::feature
H(Vector Store):::storage
I([Two-Stage Retrieval]):::feature
F ==>|"deploys via"| A
A ==>|"provisions"| B
B -->|"hosts"| E
E -->|"serves"| C
E -->|"serves"| D
G -->|"text/image/video"| C
C -->|"embeddings"| H
H -->|"top-K candidates"| D
D -->|"relevance scores"| I
A -.->|"SDK or UI"| B
classDef announced fill:#ff9900,stroke:#ec7211,color:#fff,font-weight:bold
classDef compute fill:#e3f2fd,stroke:#1565c0,color:#1565c0
classDef storage fill:#e8f5e9,stroke:#2e7d32,color:#2e7d32
classDef feature fill:#fff3e0,stroke:#e65100,color:#e65100
classDef external fill:#f5f5f5,stroke:#616161,color:#616161
What's New
AWS has made Qwen3-VL-Embedding-2B and Qwen3-Reranker-4B available in Amazon SageMaker JumpStart, bringing Qwen's latest retrieval-focused foundation models to AWS infrastructure. The embedding model handles multimodal inputs (text, images, screenshots, and video) to generate semantically rich vectors, while the reranker scores query-document pairs to refine retrieval precision. Together, they form a two-stage retrieval pipeline deployable with just a few clicks or via the SageMaker Python SDK.
How It Works
- Qwen3-VL-Embedding-2B (recall stage): Accepts text, images, screenshots, videos, and mixed-modality inputs, encoding them into a shared semantic vector space for efficient approximate nearest-neighbor search across modalities.
- Qwen3-Reranker-4B (refinement stage): Takes a (query, document) pair and outputs a scalar relevance score, allowing the pipeline to re-order the top-K candidates retrieved by the embedding model for higher precision.
- Two-stage pipeline pattern: The embedding model performs fast, broad recall over a large corpus; the reranker then applies deeper cross-attention reasoning to the shortlisted results, trading compute for accuracy only where it matters.
- User-defined instructions: The reranker accepts task-specific natural-language instructions at inference time, letting operators tune scoring behavior for specific domains, languages, or retrieval scenarios without retraining.
- Deployment via SageMaker JumpStart: Models can be deployed through the SageMaker Studio Models landing page (UI) or programmatically via the SageMaker Python SDK, provisioning a managed real-time inference endpoint on AWS-managed infrastructure.
- Language coverage: The embedding model supports 30+ languages; the reranker supports 100+ languages, including code as a first-class retrieval target.
Why It's Important
- Multimodal RAG becomes practical: Most production retrieval-augmented generation (RAG) systems today are text-only; a vision-language embedding model that natively handles images, screenshots, and video unlocks document-heavy and media-rich enterprise use cases without custom preprocessing pipelines.
- Precision-recall balance in one stack: Combining a fast embedding model with a high-accuracy reranker is the industry-standard approach to scalable search; having both available in JumpStart removes the need to source, integrate, and host them separately.
- Reduced integration overhead: Deploying through JumpStart abstracts instance selection, container management, and model serving configuration, letting teams focus on application logic rather than MLOps plumbing.
- Broad language and domain coverage: Support for 100+ languages in the reranker and code retrieval as an explicit task makes these models viable for global enterprise deployments and developer tooling (e.g., code search, documentation retrieval).
- Instruction-tunable reranking: The ability to inject task-specific instructions at inference time means a single deployed model can serve multiple downstream use cases without fine-tuning or separate endpoints.
How It's Different
- Vision-language embedding vs. text-only alternatives: Unlike most embedding models available in JumpStart (e.g., Cohere Embed, BGE), Qwen3-VL-Embedding-2B natively encodes images, screenshots, and video alongside text in a unified vector space, eliminating the need for separate vision encoders.
- Instruction-conditioned reranking: Many rerankers (e.g., cross-encoder models based on BERT/RoBERTa) are fixed-function; Qwen3-Reranker-4B accepts free-text instructions to steer scoring, offering runtime flexibility without model swaps.
- Scale of language support: At 100+ languages, the reranker significantly exceeds the coverage of most open-source cross-encoders, which typically focus on English or a handful of high-resource languages.
- Integrated two-model retrieval pattern in JumpStart: AWS is explicitly positioning these two models as a complementary pair, which is uncommon in the JumpStart catalog where models are typically listed independently without pipeline guidance.
- Parameter efficiency: At 2B (embedding) and 4B (reranker) parameters, both models are sized for cost-effective inference on mid-tier GPU instances while still delivering competitive benchmark performance, unlike larger generalist LLMs repurposed for retrieval.
When to Prefer It
- Multimodal document corpora: When your knowledge base contains PDFs with embedded images, product catalogs, slide decks, or video transcripts alongside text, and you need a single embedding space for cross-modal search.
- High-precision enterprise search: When recall from a vector database is insufficient on its own and you need a reranking stage to surface the most relevant results before passing context to an LLM in a RAG pipeline.
- Multilingual or global deployments: When your user base spans many languages and you need a reranker that can score relevance across 100+ languages without language-specific model variants.
- Code and developer tooling: When building code search, documentation retrieval, or developer assistant features that require retrieval across both natural language queries and code snippets.
- Rapid prototyping on AWS: When you want to stand up a production-grade retrieval pipeline quickly without managing custom model containers, and your team is already invested in the SageMaker ecosystem.
- Instruction-driven multi-tenant search: When a single deployment must serve multiple use cases (e.g., legal document retrieval vs. customer support FAQ) and you want to differentiate behavior via per-request instructions rather than maintaining separate endpoints.
Availability
- GA status: Both models are generally available in Amazon SageMaker JumpStart as of July 13, 2026; no preview or waitlist is indicated.
- Access methods: Available via the SageMaker Studio Models landing page (UI) and the SageMaker Python SDK for programmatic deployment.
- Regional availability: The announcement does not specify supported regions explicitly; SageMaker JumpStart model availability typically follows the SageMaker regional footprint, but customers should verify availability in their target region via the SageMaker Studio console.
- Pricing model: Charged at standard SageMaker real-time inference pricing based on the instance type and duration selected for the endpoint; no separate model licensing fee is indicated, though customers should review Qwen's open-source license terms before production use.
- License responsibility: AWS explicitly notes that customers are responsible for reviewing and complying with applicable third-party license terms before downloading or deploying JumpStart models.
- Instance requirements: Specific supported instance types are not listed in the announcement; GPU-backed instances (e.g., ml.g5 or ml.p3 family) are expected to be required given the model sizes (2B and 4B parameters).
