AWS added three foundation models to SageMaker JumpStart on August 11: NVIDIA's LocateAnything-3B, Qwen's Qwen-AgentWorld-35B-A3B, and Qwen's Qwen3.5-122B-A10B. The three cover distinct capability areas. Each carries a different infrastructure profile.
TL;DR
- -AWS added NVIDIA's LocateAnything-3B and two Qwen models to SageMaker JumpStart on August 11, covering visual grounding, agent environment simulation, and multimodal reasoning.
- -Each model targets a distinct workload type, so infrastructure, instance selection, and cost profile vary across the three.
- -Platform teams should match each model's architecture to their actual use case and check available instance configurations before assuming a large instance is required.
What Each Model Does
LocateAnything-3B is optimized for visual grounding and object localization from natural language instructions. It uses a Parallel Box Decoding (PBD) framework. PBD decodes bounding boxes and spatial points as single atomic units in one step. That design preserves geometric coherence. It also allows substantial parallel processing. AWS positions the model for Enterprise Intelligence and Physical AI applications. Supported tasks include precise object localization, dense detection, and point-based localization.
Qwen-AgentWorld-35B-A3B simulates agent environments across seven interaction domains: tool calling, search, terminal, software engineering, Android, web, and OS interaction. AWS describes it as the first language world model to cover all seven domains in a single model. It predicts the next environment state given an agent's current action and interaction history. It does this using long chain-of-thought reasoning.
The model was trained on over 10 million real-world interaction trajectories.
Qwen3.5-122B-A10B provides high-performance multimodal reasoning with production-friendly efficiency. It uses a hybrid architecture. That architecture combines Gated Delta Networks with sparse Mixture-of-Experts (MoE). MoE means only a fraction of total parameters activate per request. Of 122 billion total parameters, only 10 billion activate per token. The model uses 256 experts in its MoE layer. It carries a native context window of 262,000 tokens.
AWS notes minimal latency overhead.
Why the Architecture Differences Matter for Deployment
These are not interchangeable general-purpose models. Each targets a specific problem class. That specificity has direct infrastructure consequences.
LocateAnything-3B has 3 billion parameters. It is the lightest of the three. Teams building computer vision or Physical AI pipelines can deploy it without a large instance footprint. The single-step decoding design also matters for latency-sensitive pipelines. Both bounding box accuracy and response time count in those cases.
Qwen-AgentWorld-35B-A3B has 35 billion total parameters. Only 3 billion are active per forward pass. That ratio matters for cost planning. The active-parameter ratio means per-token compute tracks closer to a 3B model than a 35B one.
Teams evaluating agent test environments or software engineering automation should note one key point. This model covers all seven interaction domains in one deployment. That avoids maintaining separate specialized models per domain.
Qwen3.5-122B-A10B is the heaviest of the three by total parameter count. The 10 billion active parameter ceiling keeps per-token compute lower than the raw figure suggests. The 262K native context window is the practical differentiator. Agent workflows and multimodal tasks that accumulate long histories can fit more working context in a single call.
Teams running high-cost chunking or summarization steps to work around smaller context limits should evaluate whether this model removes those intermediate steps.
What Platform Teams Need to Decide
All three models are deployable through the SageMaker JumpStart model catalog in the SageMaker console, or via the SageMaker Python SDK. Deployment itself is a few clicks or SDK calls. The harder decisions come before that.
Instance selection is the first cost lever. SageMaker JumpStart supports multiple deployment configurations per model. Each configuration tunes the model for one use case and one performance goal: lowest cost, highest throughput, lowest latency, or balanced price-performance. A configuration often supports a smaller or less expensive instance than the default. Platform teams should list available configurations for each model before concluding that a large instance is required.
Control boundary is the second consideration. All three models deploy into the team's own AWS account as managed endpoints. For teams handling sensitive data, that account boundary should be confirmed against data governance requirements before production use.
License terms warrant a check before production deployment. SageMaker JumpStart documentation advises verifying license terms with the model provider when the catalog lists no license for a model. If the catalog lists no license for any of these three models, platform teams should confirm the applicable license directly with NVIDIA or Qwen before integrating that model into commercial workflows.
Finally, these three models cover distinct workload types. A team evaluating agent infrastructure should map Qwen-AgentWorld-35B-A3B against its specific interaction domains. A team building computer vision pipelines should assess LocateAnything-3B's spatial accuracy on its actual target objects. Qwen3.5-122B-A10B is the candidate to test when long context, multimodal inputs, or reasoning depth is the binding constraint. The catalog addition expands real options.
Choosing the right one requires matching the architecture to the actual workload.
Sources and supporting resources