SageMaker JumpStart adds over 100 models from 30 providers in 2026

What SageMaker JumpStart offers

SageMaker JumpStart is a machine learning hub within Amazon SageMaker AI. It provides a curated catalog of AI models, solution templates and example notebooks, and AWS says it shortens the path from model selection to deployment. The service covers three main use cases: discovering and evaluating models, deploying them with one click to a SageMaker real-time inference endpoint, and customizing them.

For customization, users can fine-tune models from the console or SDK with techniques such as low-rank adaptation (LoRA), supervised fine-tuning (SFT), direct preference optimization (DPO) and reinforcement learning with verifiable rewards (RLVR). Access is possible through the SageMaker Studio console, the SageMaker Python SDK, or directly from Hugging Face via a recently launched deep link integration.

Which providers contributed the most models

AWS says the growth was driven by the accelerating pace of open source model releases and growing customer demand for choice. The breakdown by provider:

• Alibaba (Qwen): 29 models, covering reasoning, mixture-of-experts, agentic workflows, code, speech recognition, text-to-speech, embeddings and reranking.
• NVIDIA: 17 models, including the Cosmos3 world foundation model family and several Nemotron models.
• Mistral: 7 models, including the Ministral-3 family and Voxtral-Mini-4B-Realtime.
• Google: 6 Gemma 4 variants.
• IBM: 4 models, including Granite 4.0 variants and granite-speech-4.1-2b.
• Cohere: 4 models, including Command A and Rerank versions.
• Others: 33 models from providers such as MiniMax, Zhipu AI, DeepSeek, Moonshot AI, Allen AI, Microsoft, Meta, Black Forest Labs, JetBrains, OpenBMB, BharatGenAI, Kakao and OpenFold.

AWS notes that the catalog now spans 30 providers, up from what it calls a more concentrated set in prior years, and argues that reducing concentration risk matters to enterprise customers.

Focus areas: multimodal, compact and specialized models

AWS describes several priorities. For NVIDIA, it highlights the Cosmos3 models in Super, Edge and Nano tiers, which it says combine video generation, physics simulation and action planning for physical AI uses such as robotics, autonomous systems and digital twins.

On multimodal workloads, the company points to NVIDIA Nemotron 3 Nano Omni, which processes text, vision and speech in one model, as well as Gemma 4, granite-speech-4.1-2b and Voxtral-Mini. AWS says these additions let customers avoid chaining separate preprocessing models.

For creative AI, the catalog gained FLUX.2 Small Decoder and FLUX.2-klein-base-4B from Black Forest Labs, Meta Muse-Glimmer-30B and NVIDIA Cosmos3 for video generation.

AWS also onboarded efficient and quantized models for customers focused on inference costs. Examples include Zhipu AI GLM-5.2 FP8, JetBrains Mellum2-12B with only 2.5B active parameters, and OpenBMB MiniCPM5-1B with 1 billion parameters. According to AWS, Microsoft Phi-4-mini-instruct and IBM Granite 4.0 h-tiny show that compact models can serve production workloads on instance types like ml.g5.xlarge.

Purpose-built models round out the list: Qwen-AgentWorld-35B-A3B for autonomous workflows with native tool-calling, GLM-OCR, LightOnOCR and DeepSeek-OCR for document text extraction, Qwen3-Reranker-4B and Cohere Rerank v4.0 for retrieval precision in RAG pipelines, and OpenFold3 for protein structure prediction.

Platform improvements: Model Hub and deployment configurations

AWS also shipped changes to reduce friction. Previously, finding JumpStart models meant going into SageMaker Studio and locating the JumpStart section in the left navigation. The new Model Hub surfaces the full catalog on the SageMaker AI console landing page. Users can browse by category and filter by input/output modality, publisher, model size, license and framework, view sample notebooks, and open models in Studio with one click.

For deployment, AWS says choosing an instance type, serving container and optimization settings typically required manual benchmarking across dozens of combinations. JumpStart now offers named deployment configurations, including a Balanced option that trades off cost, latency and throughput evenly for general-purpose workloads. The source text breaks off while listing the further options, which earlier in the article are described as low cost, low latency, throughput or balanced.

What it means for users

For teams building on SageMaker, the update means a broader choice of open models from a wider set of providers, with the same one-click deployment path across language, vision, speech and image models. The mix of compact and quantized models gives those watching inference costs more options, while the Model Hub and preset configurations aim to shorten the time between exploring a model and running it. These benefits reflect AWS's own account of the catalog and features.