AWS adds three open-weight DeepSeek models to SageMaker JumpStart

AWS has made three open-weight DeepSeek models available through SageMaker JumpStart, including a model AWS says supports a one-million-token context window and another designed for agentic workloads. For AWS customers, the practical change is a new route to deploy them through the service, rather than a claim that their advertised performance has been independently proved.

AWS AI Watch analysis

What happened

AWS says DeepSeek-V4-Pro, DeepSeek-V4-Flash and DeepSeek-V4.1-Flash are now available in SageMaker JumpStart. The models are released under the MIT licence, according to AWS, which says customers can deploy them through the SageMaker model catalogue or Python SDK. AWS’s announcement describes three different options: V4-Pro for demanding reasoning and generation, V4-Flash for lower-compute deployments, and V4.1-Flash for agentic work, coding and vision.

The figures AWS gives are substantial. It describes V4-Pro as a 1.6-trillion-parameter mixture-of-experts model activating 49 billion parameters per token, with a one-million-token context window. V4-Flash is listed at 284 billion parameters, activating 13 billion per token, and has the same context-window length. AWS says V4.1-Flash has a 552-billion-parameter backbone, activates 8 billion parameters during prefill and 16 billion during decoding, and supports up to 384,000 output tokens.

Why it matters

JumpStart gives AWS customers a deployment option for all three models within its managed machine-learning service. The choice is not simply “largest is best”: the model descriptions point to different trade-offs in scale, compute and intended workload. Those deployment details may help teams decide which one is worth testing against their own tasks and infrastructure.

AWS also makes performance and efficiency claims, including that the models deliver frontier-class results and that V4.1-Flash’s caching reduces its global KV-cache footprint. Those are the company’s descriptions, not independent benchmark findings. A very large context window is a specification, not a guarantee that every task will make good use of it.

Our read

This is a meaningful addition to the menu for AWS-based AI teams, particularly those evaluating open-weight models without building a deployment route from scratch. The useful next step is a workload-specific trial: compare quality, latency and cost, rather than choosing by parameter count or the particularly grand number attached to the context window.

What to watch

• Which regions and deployment configurations AWS supports for each model.
• How the models perform on independent benchmarks and real customer workloads.
• Whether the advertised context and output limits are available in practical deployments.

Discussion spark: When choosing an open-weight model for production, would you prioritise the largest context window, lower compute needs or performance on your own workload?

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