AWS AI Updates: August 28, 2026

1. NVIDIA Cosmos3 World Models Land on SageMaker JumpStart

Amazon SageMaker JumpStart now offers NVIDIA’s Cosmos3 family of open omnimodal world models for physical AI, spanning robotics, autonomous vehicles, and vision systems. The lineup covers three tiers: Cosmos3-Edge (4B parameters, tuned for on-device robot control at 32 actions per inference at 15 Hz on Jetson Thor), Cosmos3-Nano (16B parameters for physics-aware world generation across text, image, video, audio, and action trajectories), and Cosmos3-Super (64B parameters using a Mixture-of-Transformers architecture for high-fidelity simulation and synthetic data generation). All three deploy from the JumpStart catalog with minimal configuration through the SageMaker console or Python SDK, giving teams building simulation and policy-learning pipelines a managed path to frontier world models. Source

2. Muse-Glimmer-30B and Qwen 3.8-27B Added to SageMaker JumpStart

Amazon SageMaker JumpStart added two agentic foundation models: Meta’s Muse-Glimmer-30B and Alibaba’s Qwen 3.8-27B. Muse-Glimmer-30B is a 30B dense model with a roughly 1.8B ViT-G/14 perception encoder, a 131K-plus token context window, selectable reasoning strength, and an Apache 2.0 license, targeting multi-step agentic tasks, tool use, and failure recovery. Qwen 3.8-27B is a native vision-language model with a 262K context window extendable to about 1M via YaRN scaling, reports 61.7 on SWE-bench Pro, and runs at roughly 17GB quantized for coding and multimodal reasoning across text, images, and video. Both are deployable in a few clicks from the JumpStart catalog via the SageMaker console or Python SDK. Source