Amazon SageMaker JumpStart models expand with 100+ new additions
Amazon SageMaker JumpStart models have grown by over 100 new additions from 30 providers, including NVIDIA and Alibaba, in 2026.
Amazon SageMaker JumpStart models have seen a significant expansion in 2026, with more than 100 new models from 30 different providers added to the catalog between January and September. This growth provides developers and businesses with a much wider selection of foundation models to power various artificial intelligence applications.
What new models are in SageMaker JumpStart?
Between January and September 2026, the platform onboarded over 100 new models, reflecting a major increase in available architectures and capabilities. The expansion includes contributions from a diverse range of providers:
- Alibaba (Qwen): 29 models covering reasoning, large-scale mixture-of-experts (MoE), agentic workflows, code generation, speech recognition, text-to-speech, embeddings, and reranking.
- NVIDIA: 17 models, including the Cosmos3 world foundation model family (Super, Edge, and Nano tiers), which support video generation, physics simulation, and action planning. Other models include the Nemotron and MSA-Search families.
- Mistral: 7 models, including the Ministral-3 family and the Voxtral-Mini-4B-Realtime for speech.
- Google: 6 variants of Gemma 4, including different sizes and quantized versions.
- IBM: 4 models, including Granite 4.0 and granite-speech-4.1-2b for multilingual automatic speech recognition.
- Cohere: 4 models, including Command A and various Rerank versions for retrieval augmentation.
- Other Providers: 33 models from companies such as Microsoft, Meta, DeepSeek, Zhipu AI, and Moonshot AI.
Why is the model expansion important?
The addition of these models matters because it gives users more flexibility to choose the right tool for a specific task. Instead of being limited to a few options, developers can now select models based on specific needs like low latency, high throughput, or specialized capabilities such as code generation or video simulation.
By providing a curated catalog, SageMaker JumpStart reduces the time required to move from selecting a model to actually deploying it in a production environment. This helps businesses build and scale AI applications more efficiently.
How does SageMaker JumpStart work?
SageMaker JumpStart acts as a central hub within SageMaker AI that simplifies the entire machine learning lifecycle, from finding a model to customizing it for your specific data.
Model Discovery and Evaluation: Users can browse hundreds of foundation models. Each model includes a "model card" containing technical architecture details, performance benchmarks, licensing information, and sample notebooks to help with evaluation.
One-Click Deployment: Once a model is selected, it can be deployed to a SageMaker real-time inference endpoint. The system uses pre-configured settings optimized for different workloads, such as low cost, low latency, or balanced performance, removing the need for manual container or instance configuration.
Model Customization: Users can fine-tune models directly through the console or the SageMaker Python SDK. This customization uses advanced techniques such as:
- Low-Rank Adaptation (LoRA)
- Supervised Fine-Tuning (SFT)
- Direct Preference Optimization (DPO)
- Reinforcement Learning with Verifiable Rewards (RLVR)
For those looking to minimize infrastructure management, serverless customization options are also available.
How can users access these models?
There are three primary ways to interact with the catalog:
- Through the SageMaker Studio console via the Models menu.
- Programmatically using the SageMaker Python SDK.
- Directly from Hugging Face through a recently launched deep link integration.
Source: aws.amazon.com