AWS says SageMaker JumpStart added more than 100 models from 30 providers

AWS says SageMaker JumpStart added more than 100 models from 30 providers between January and September 2026, alongside new tools for discovering and deploying them. The expansion gives customers a wider choice of models and a more direct route from selection to an AWS endpoint.

AWS AI Watch analysis

What happened

In its 2026 year-to-date review, AWS says JumpStart’s catalogue now spans models from providers including Alibaba, NVIDIA, Google, Mistral and Cohere. Alibaba’s Qwen family contributed 29 models, while NVIDIA contributed 17.

AWS also describes changes to the service: a Model Hub on the SageMaker AI console, pre-configured deployment choices for cost, latency or throughput, and a Hugging Face link that can take users into a configured SageMaker Studio environment for supported models. The review also says customers can fine-tune models through serverless customisation options.

Why it matters

For teams choosing models, a broader catalogue is useful only if they can compare and try options without turning every experiment into a miniature infrastructure project. AWS says model cards include details such as benchmarks, licences and sample notebooks; its named deployment profiles are intended to reduce the manual work of configuring and testing endpoints.

The catalogue also covers more than chat. AWS highlights models for vision, speech, document understanding, image and video generation, retrieval and scientific computing. That range could help teams select a specialist model for a task rather than ask one general-purpose model to do everything, which is an appealing way to avoid giving a Swiss Army knife the keys to the workshop.

Our read

This is a meaningful expansion of AWS’s model menu, with practical changes to browsing and deployment that may make experimentation less fiddly. It is AWS’s account of its own service, not an independent comparison of model quality or deployment costs. The useful next step for prospective users is to check the model’s licence, regional availability, hardware needs and performance on their own workload before committing.

What to watch

• Whether the new catalogue and filters make it easier to find suitable models.
• How the pre-configured profiles perform on real workloads, including their cost and latency.
• Which newly listed models gain broader regional availability and continued support.

Discussion spark: When choosing a model for a production workload, should teams prioritise having more providers to choose from, or a smaller catalogue with stronger evidence about cost and performance?

not affiliated with or endorsed by Amazon Web Services (AWS)