Amazon says its SageMaker JumpStart machine learning hub onboarded more than 100 models from 30 providers between January and September 2026. The catalog additions include models from Alibaba’s Qwen, NVIDIA, Mistral, Google, IBM, Cohere and other providers, spanning language, vision, speech, image, video, retrieval and scientific applications.

According to Amazon, the models include NVIDIA’s Cosmos3 and Nemotron families, Google’s Gemma 4 variants, IBM Granite models, Mistral’s Ministral and Voxtral models, and specialized systems such as Qwen3-Reranker, Qwen3-ASR, Qwen3-TTS and OpenFold3. Amazon also highlighted compact and quantized models, including Microsoft Phi-4-mini-instruct and IBM Granite 4.0 h-tiny.

The company said JumpStart’s 2026 enhancements include a Model Hub experience on the SageMaker AI console. Users can browse the catalog and filter models by modality, publisher, size, license and framework, while accessing sample notebooks and SageMaker Studio. Amazon also introduced named deployment profiles for balanced, throughput-, latency- and cost-optimized configurations, including setups for AWS Inferentia and Trainium instances.

Amazon said a new Hugging Face deep-link integration allows users to choose options to deploy or customize supported models on SageMaker AI. The company said the process opens a configured Studio environment with the model and permissions prepared. JumpStart remains available through SageMaker Studio and the SageMaker Python SDK.