Daily News · 3 min read

AWS AI Updates: August 12, 2026

aws bedrock sagemaker generative-ai

1. Bedrock Extends Per-Principal Cost Allocation to the bedrock-mantle Endpoint

AWS. Amazon Bedrock now supports cost allocation by IAM principal for the bedrock-mantle endpoint, extending a capability previously limited to the bedrock-runtime endpoint. Teams can tag IAM users and roles with attributes like team, project, or cost center, activate them as cost allocation tags, and attribute inference spend through AWS Cost Explorer or Cost and Usage Report 2.0 with caller-identity data. The result is precise chargeback and budget attribution for generative AI inference across users, teams, and applications in production. Source

2. Visual Grounding, Agent Simulation, and Sparse MoE Models Land on SageMaker JumpStart

AWS. Amazon added three foundation models to SageMaker JumpStart targeting distinct workloads. LocateAnything-3B uses a Parallel Box Decoding framework to emit bounding boxes and points as atomic units in a single step for object localization, and Qwen-AgentWorld-35B-A3B is a language world model trained on more than 10 million interaction trajectories spanning tool calling, terminal, software engineering, Android, web, and OS domains. Qwen3.5-122B-A10B is a 122B-parameter sparse Mixture-of-Experts model with 256 experts and 10B active parameters per token, pairing Gated Delta Networks with a native 262K-token context window, and all three deploy from the JumpStart catalog or the SageMaker Python SDK. Source

3. NVIDIA Nemotron 3.5 Lightning Arrives on SageMaker JumpStart

AWS. NVIDIA’s Nemotron 3.5 Lightning, a hybrid Mixture-of-Experts model with 30B total parameters and 3B active per forward pass, is now deployable from SageMaker JumpStart. AWS cites roughly 410 tokens per second, up to 4x the throughput and 30% faster task completion versus comparable models, plus a 1M-token context window via DFlash speculative decoding, aimed at persistent agents and high-throughput automation such as financial document processing, cybersecurity triage, and telecom operations. Trained on open datasets, the model lets enterprises customize and retain ownership, and deploys in a few clicks from the JumpStart catalog or the SageMaker Python SDK. Source

4. AWS Glue Adds One-Click Navigation into SageMaker Unified Studio

AWS. The AWS Glue console now offers one-click navigation into Amazon SageMaker Unified Studio, letting data engineers move from browsing catalog tables to querying data, running data quality checks, and building pipelines using the same IAM role without switching consoles. An inline permissions panel lets new users customize IAM access in context, cutting setup steps. The integration streamlines the path from data cataloging into SageMaker’s analytics and ML capabilities and is available in all regions that support SageMaker Unified Studio. Source