Amazon Web Services announced that three DeepSeek language models are now listed in SageMaker JumpStart. The models – DeepSeek‑V4‑Pro, DeepSeek‑V4‑Flash and DeepSeek‑V4.1‑Flash – are open‑weight, meaning their code and weights are freely available under an MIT license. Customers can add them to their AWS accounts with a few clicks in the console or by using the SageMaker Python SDK. The announcement was made on November 12, 2026.
DeepSeek‑V4‑Pro is the largest of the three. It has 1.6 trillion parameters in a Mixture‑of‑Experts (MoE) architecture, but only 49 billion of those are active for each token. It also supports a 1 million‑token context window, which lets it read and reason over very long documents. This makes it useful for tasks like multi‑step problem solving, detailed report generation, or any application that needs maximum language capacity.
DeepSeek‑V4‑Flash is a smaller, cheaper alternative. With 284 billion total parameters and 13 billion active per token, it keeps the same 1 million‑token window but uses far less compute. AWS positions it for latency‑sensitive production services where cost matters, such as chatbots or real‑time summarisation.
The newest entry, DeepSeek‑V4.1‑Flash, adds multimodal support and a novel Causal Encoder‑Decoder design. Its backbone holds 552 billion parameters, yet only 8 billion are used during the initial prefill and 16 billion while generating text. It can output up to 384 k tokens and uses Compressed Sparse Attention 2 (CSA2) plus FP4 KV caching to shrink memory usage to about a quarter of V4‑Flash. These tricks aim at agentic workloads that combine text and vision, such as code assistants that also read screenshots.
Why add these models now?
AWS expands its catalog of foundation models to give developers more choices between raw power and efficiency. By offering both a flagship‑scale model (V4‑Pro) and two cost‑focused variants (V4‑Flash and V4.1‑Flash), customers can pick the right balance for their budgets and latency requirements without leaving the AWS ecosystem. The open‑weight licence also lets smaller teams fine‑tune the models or integrate them into proprietary pipelines.
Why it matters
Developers can now run large‑scale language models on AWS without building their own infrastructure, choosing a version that matches their cost or speed needs. The open‑weight licence also means smaller firms can customize the models for niche applications, though they still need AWS credits or a paid account to deploy them.
