AWS Adds Three Specialized AI Models to SageMaker JumpStart

Amazon SageMaker JumpStart now hosts three distinct models built for very different jobs: pinpointing objects in images, navigating software and web interfaces, and handling long-context reasoning at scale. AWS made all three available in January 2026, giving developers a single deployment path for capabilities that previously required stitching together separate research releases.

The first, NVIDIA's LocateAnything-3B, is a visual grounding model that identifies and localizes objects within images. It relies on a technique called Parallel Box Decoding (PBD), which lets the model predict multiple bounding boxes simultaneously rather than one at a time, a design choice aimed at speeding up object-detection workflows.

The second addition, Qwen-AgentWorld-35B-A3B from Alibaba's Qwen team, is built for AI agents that need to act rather than just answer questions. It was trained on more than 10 million real-world interaction trajectories spanning seven domains: tool calling, search, terminal use, software engineering, Android, web browsing, and general OS interaction. That breadth positions it as a foundation for agents that operate across multiple digital environments instead of a single narrow task.

The third model, Qwen3.5-122B-A10B, targets long-context, compute-efficient inference. It carries 122 billion total parameters but activates only 10 billion per token, a Mixture-of-Experts design that keeps inference costs closer to a much smaller model while retaining a large parameter pool.

That efficiency comes from a hybrid architecture combining Gated Delta Networks with a sparse Mixture-of-Experts layer built from 256 experts. Paired with a native 262K-token context window, the model is positioned for tasks like long-document analysis, extended agent sessions, or codebase-scale reasoning where context length is often the bottleneck.

Why bundling matters

Individually, none of these three releases is unprecedented: visual grounding, agentic training corpora, and sparse MoE architectures with long context have all appeared elsewhere in 2025. What AWS is offering is operational rather than technical, a managed on-ramp through SageMaker JumpStart that removes the setup overhead of self-hosting each model independently.

For teams building multimodal or agentic pipelines, that means LocateAnything-3B can handle the perception layer, Qwen-AgentWorld-35B-A3B can drive interaction and action, and Qwen3.5-122B-A10B can serve as the long-context reasoning backbone, all provisioned from the same console. The announcement itself is light on benchmark comparisons or pricing detail, so how these models perform against incumbents like GPT-4o-class agents or existing SageMaker offerings remains to be demonstrated in practice rather than in AWS's release notes.