What's New

AWS has added three new models to Amazon SageMaker JumpStart: paraphrase-multilingual-MiniLM-L12-v2 (a lightweight multilingual sentence embedding model from Sentence Transformers), Microsoft Table Transformer Detection (a DETR-based table detection model for unstructured documents), and Bielik-11B-v3.0-Instruct (an 11B-parameter generative language model optimized for Polish and European languages). All three models are now available for one-click deployment directly from SageMaker Studio or via the SageMaker Python SDK. This expansion broadens JumpStart's model catalog with specialized capabilities spanning multilingual NLP, document understanding, and European-language generative AI.

How It Works

• Each model can be deployed through SageMaker JumpStart by navigating to the Models section of SageMaker Studio or by using the SageMaker Python SDK, which provisions the necessary inference infrastructure automatically.
• Paraphrase-multilingual-MiniLM-L12-v2 encodes input text into 384-dimensional dense vectors using a transformer architecture fine-tuned for semantic similarity across 50+ languages, enabling cosine similarity comparisons for search and clustering tasks.
• Table Transformer Detection applies a DETR (DEtection TRansformer) architecture trained on the PubTables-1M dataset to perform bounding-box-level table localization within document images or PDFs, outputting detected table regions suitable for downstream extraction pipelines.
• Bielik-11B-v3.0-Instruct is a 11-billion-parameter instruction-tuned causal language model trained on multilingual corpora covering 32 European languages, designed to handle dialogue, reasoning, and tool-use tasks with particular strength in Polish-language contexts.

Why It's Important

• These additions address three distinct but underserved use cases in enterprise AI workflows.
• The multilingual MiniLM model enables cost-efficient semantic search and document clustering across language boundaries without requiring separate language-specific models, which is critical for globally operating organizations.
• Table Transformer Detection fills a significant gap in document digitization pipelines—reliably locating tables in PDFs and scanned images is a foundational step for automating data extraction from financial reports, research papers, and regulatory documents.
• Bielik-11B-v3.0-Instruct is particularly significant for European enterprises and public sector organizations that require strong Polish and broader European language support, a segment historically underserved by predominantly English-centric LLMs.
• Together, they lower the barrier to deploying specialized, production-grade AI capabilities without requiring customers to source, containerize, or optimize these models independently.

How It's Different

• Prior to these additions, SageMaker JumpStart's multilingual embedding options were largely dominated by larger or more general-purpose models, whereas paraphrase-multilingual-MiniLM-L12-v2 offers a deliberately lightweight footprint (384 dimensions, 12 transformer layers) that trades some representational capacity for significantly lower inference cost and latency.
• Table Transformer Detection is purpose-built for table detection specifically—unlike general-purpose object detection models (e.g., YOLO or Detectron2 variants) that require fine-tuning on document data, this model is pre-trained on PubTables-1M and ready for document-domain use out of the box.
• Bielik-11B-v3.0-Instruct differentiates itself from other JumpStart generative models by its explicit focus on Polish and Central/Eastern European languages, offering a level of linguistic depth for those languages that general multilingual models like Mistral or LLaMA variants typically do not match.
• The JumpStart deployment mechanism itself differentiates all three from self-managed deployments by abstracting infrastructure provisioning, model artifact management, and endpoint configuration.

When to Prefer It

• Choose paraphrase-multilingual-MiniLM-L12-v2 when you need fast, low-cost semantic similarity or embedding generation across multiple languages and can accept a compact vector representation (384 dimensions); it is ideal for high-throughput retrieval-augmented generation (RAG) pipelines, multilingual FAQ matching, or document clustering where inference cost is a primary constraint.
• Prefer Table Transformer Detection when your pipeline involves structured data extraction from PDFs, scanned documents, or research literature and you need reliable table localization as a preprocessing step before OCR or data parsing—particularly in financial services, life sciences, or legal document workflows.
• Select Bielik-11B-v3.0-Instruct when your application requires high-quality Polish-language generation, reasoning, or dialogue, or when serving users across Central and Eastern Europe where other LLMs exhibit degraded performance; it is also well-suited for STEM and mathematical reasoning tasks in a multilingual European context.
• If your use case demands English-only or broader global language coverage with maximum generative capability, other JumpStart models such as Llama or Mistral variants may be more appropriate.

Availability

• All three models are generally available (GA) in Amazon SageMaker JumpStart as of April 29, 2026.
• Deployment is accessible through SageMaker Studio's Models section or the SageMaker Python SDK within any AWS account that has SageMaker enabled.
• The announcement does not specify regional restrictions, but SageMaker JumpStart availability is generally tied to regions where SageMaker Studio and the underlying EC2 instance families (e.g., ml.g4dn, ml.g5, ml.p3) are supported; customers should verify instance availability in their target region before deployment.
• No explicit mention is made of VPC-only deployment constraints, model fine-tuning support, or batch transform compatibility for these specific models, so consulting the SageMaker JumpStart documentation is recommended for detailed configuration options and any service quota considerations.