AI Lab

Amazon News

Official Amazon announcements — model releases, product launches and research, each one summarized with every source covering it, by AIBriefs. RSS

LaunchDevelopers1 source

AWS adds managed NVIDIA Dynamo to SageMaker HyperPod inference

AWS now manages NVIDIA Dynamo on SageMaker HyperPod, splitting prompt processing and token generation across GPUs. AWS cites GLM-5.3's FP8 weights exceeding the 640 GB of GPU memory on a single ml.p5.48xlarge instance before KV cache and runtime.

LaunchDevelopers1 source

SageMaker AI real-time endpoints add KV cache-aware routing

The new routing strategy reads cache state reported by the model container and sends each request to an instance holding a long cached prompt match that isn't too busy. It targets real-time endpoints on Amazon SageMaker AI.

How-ToDevelopers1 source

AWS details secure OAuth patterns for Snowflake Cortex AI on Bedrock AgentCore

AWS published a technical how-to for extending enterprise authentication and authorization controls into the agentic layer when an AI agent on Amazon Bedrock AgentCore calls Snowflake Cortex AI. It targets the common workaround of broad service-account permissions or custom-built auth integrations that add operational overhead.

LaunchDevelopers1 source

Amazon SageMaker adds AI data preparation capabilities

AWS announced AI data preparation capabilities in Amazon SageMaker for building training and evaluation datasets. AWS says fine-tuning open-weight models, running reinforcement learning, and building evaluation suites all depend on high-quality data.

LaunchBusiness1 source

AWS expands Forward Deployed Engineering to 15 countries

AWS Forward Deployed Engineering is now available in 15 countries, including the U.S., Mexico, Brazil, Canada, Australia, Singapore, India, South Korea, Japan, the UK, Germany, the Netherlands, France, Sweden and Saudi Arabia. It serves both commercial and public sector customers.

How-ToDevelopers1 source

AWS shows how to build WhatsApp AI agents on Bedrock AgentCore

AWS published a technical walkthrough for building WhatsApp AI agents on Amazon Bedrock AgentCore, aimed at developers in Latin America, India, and Southeast Asia where WhatsApp is a dominant business channel. AWS calls WhatsApp a genuinely difficult surface to build on.

How-ToDevelopers1 source

OpenTSLM time-series model now on Amazon SageMaker JumpStart

AWS published a technical walkthrough for running OpenTSLM, a model for reasoning over time-series data, via SageMaker JumpStart. The post frames it around industrial use cases like preventing production stoppages and diagnosing vehicle faults.

How-ToPolicy1 source

AWS publishes phased roadmap for responsible agentic AI

AWS's post targets platform and application teams building agentic AI platforms that govern the applications, tools, skills, models, and storage agents use. It argues generative AI's nondeterministic nature adds complexity prior software practices didn't anticipate.

How-ToAI Agents1 source

AWS maps multi-agent design patterns and anti-patterns

AWS argues multi-agent failures are rarely the model's fault — coordination is a distributed systems problem, and many teams lack distributed systems instincts. The post catalogs design patterns, anti-patterns, and lessons drawn from manufacturing.

How-ToDevelopers1 source

Amazon Bedrock adds per-tenant answer customization

Retrieval instructions in Amazon Bedrock Managed Knowledge Base let teams set voice and answer behavior per tenant, so shared product and policy documents return different answers to different audiences.

LaunchScience1 source

OpenFold3 biomolecular structure model lands on SageMaker JumpStart

OpenFold3-preview, an open-source biomolecular structure prediction model from the OpenFold Consortium and Columbia's AlQuraishi Lab, can now be deployed from Amazon SageMaker JumpStart with a few lines of code into a customer's own AWS environment.

How-ToDevelopers1 source

AWS shows SageMaker HyperPod Data Capture for model improvement

AWS's technical walkthrough covers capturing production inference data on SageMaker HyperPod to detect drift, surface failure patterns, and feed real inputs back into retraining. It targets gaps no pre-curated dataset covers, such as domain jargon, unexpected formats, and shifting user behavior.

LaunchDevelopers2 sources

SGLang-Neuron brings SGLang serving to AWS Trainium

SGLang-Neuron is publicly available via an AWS Deep Learning Container, released by RadixArk with Annapurna Labs, for serving models and generating RL rollouts on Trainium. SGLang already supports NVIDIA, AMD, Intel and Google chips, so teams can switch hardware without changing software.

How-ToAI Agents1 source

Natera built SWARM lab-ops agent on Amazon Bedrock AgentCore

Natera's lab engineering team, led by Ariel Jirau, built SWARM (Systematic Workflow Analytics and Resolutions Manager) with AWS to keep cell-free DNA testing instruments running, since equipment downtime disrupts sample processing.

How-ToDevelopers1 source

AWS blog: four Bedrock inference settings that drive cost

Cache duration, service tier, reasoning effort, and maxTokens can change how many tokens Bedrock processes and the rate applied to them, per AWS's guide. The post argues bills keep climbing even after prompt and application-logic tuning.

How-ToDevelopers1 source

AWS guide covers cost-per-task and agent ROI on Bedrock

Part 2 of an AWS series on controlling Amazon Bedrock spend, following part 1's model of IAM principal-tag attribution, Cost and Usage Report 2.0, model invocation logging, and three-tier enforcement.

How-ToDevelopers1 source

AWS shows how to build a self-service LLM gateway with Bedrock and LiteLLM

AWS's technical how-to pairs Amazon Bedrock foundation models with LiteLLM so developers get quick access to models while platform teams keep visibility and usage control. The setup targets latency and error monitoring plus cost oversight for internal AI coding assistants.

How-ToAI Agents1 source

AWS shows how to build autonomous AI agents with Bedrock and Spring AI

AWS tutorial targets Java teams moving agents from prototype to production, pairing Amazon Bedrock with Spring AI on existing JVM tooling and CI/CD pipelines. It argues agent frameworks alone don't cover scalability, governance, security, or operational concerns.

How-ToDevelopers1 source

OpenAI Codex now runs on Amazon Bedrock

AWS published a technical how-to on running OpenAI Codex through Amazon Bedrock so platform teams avoid standing up a second model-access, identity, and operations stack. The post covers using Codex to explore unfamiliar code, make changes, and handle repetitive engineering work.

LaunchAI Agents3 sources

AWS launches Web Search API on Amazon Bedrock AgentCore

The new API gives agents access to information newer than their model's training data, covering recent events, product releases, prices, and regulations. It replaces the previous route of delivering web search to agents through AgentCore Gateway.

How-ToDevelopers1 source

AWS recaps October 2026 AI builder releases on Bedrock

AWS's monthly roundup covers October 2026 additions across Amazon Bedrock and Bedrock AgentCore, framed around model choice, production trust for agents, and cost visibility. The post is a recap of already-shipped features rather than a new launch.

How-ToDevelopers1 source

AWS shows push-based sync for Agent Registry MCP updates

AWS published a how-to for automating AWS Agent Registry record updates whenever a published MCP server adds a tool or changes its schema, avoiding stale metadata for consumers. The registry is a capability of Amazon Bedrock AgentCore.

How-ToAI Agents1 source

DXC Technology builds multi-model AI substrate on Amazon Bedrock

AWS customer post co-written with DXC's Nick Tamburro describes a multi-model substrate on Bedrock for agents acting on production servers at banks and hospitals. It cites recent public incidents where autonomous agents took destructive actions their operators never intended.

How-ToDevelopers1 source

AWS blog: reusable assets speed generative AI delivery

AWS's post argues teams waste time rebuilding RAG pipelines, document processors, and agent scaffolding instead of reusing existing components. It is tagged Intermediate (200) and Technical How-to, aimed at Amazon Bedrock builders.

How-ToAI Agents1 source

AWS details UAEF, a framework-agnostic agent evaluation system

UAEF targets teams running agents across LangGraph, LangChain, Strands Agents, and Amazon Bedrock AgentCore, whose differing execution records make cross-framework evaluation hard. It provides closed-loop evaluation that normalizes message, tool-call, and handoff structures across frameworks.

LaunchDevelopers1 source

Kimi K3.1 Pro from Moonshot AI lands on Amazon Bedrock

Kimi K3.1 Pro Highspeed is a post-trained version of Kimi K3 with the same 3T-class open-weight architecture and a 1-million-token context window, trained with substantially more post-training compute.

How-ToAI Agents1 source

AWS publishes agent evaluation playbook for Bedrock AgentCore

AWS's guide covers three evaluation approaches on AgentCore Evaluations: built-in evaluators, DeepEval, and AutoEval. It targets teams that currently spot-check outputs and watch latency instead of measuring whether agents meet business needs at scale.

How-ToAI Agents1 source

Trovata builds autonomous treasury agents on Amazon Bedrock and AgentCore

Trovata's AI-native treasury solution runs on Amazon Bedrock and AgentCore, targeting banking data fragmented across dozens of bank portals and manual reporting trapped in spreadsheets. Legacy Treasury Management Systems are cited as unable to scale for modern enterprise needs.

How-ToAI Agents1 source

TCS builds agentic test orchestrator on Amazon Bedrock AgentCore

TCS uses Amazon Bedrock AgentCore to automate large-scale testing for banking, financial services, and insurance clients, where release delays stem from testing rather than application code. The post is co-written by TCS's Varun Valsan.

How-ToDevelopers1 source

AWS publishes Bedrock tokenomics data lake guide

AWS's walkthrough targets teams running thousands of model invocations a day across multiple accounts, where the aggregate bill is visible but per-request detail is not. It builds an observability data lake for Amazon Bedrock workloads.

How-ToAI Agents1 source

AWS Quick encodes repeatable tasks as agents and skills

AWS blog post explains how AWS Quick lets users encode repeatable tasks — weekly status updates, release notes, project management, ticket creation — as agents and skills instead of re-explaining them each time.

How-ToDevelopers1 source

Amazon Bedrock Managed Knowledge Base adds agentic retrieval

AWS walks through single-shot versus agentic retrieval patterns on Amazon Bedrock Managed Knowledge Base, using a banking assistant that answers document-checklist questions and then compares requirements across sole proprietorships, LLCs, and partnerships.

LaunchAI Models1 source

NVIDIA Nemotron 3.5 Super VL lands in Amazon SageMaker JumpStart

Super VL is the vision-language model in the Nemotron 3.5 family, a 120B-parameter mixture-of-experts that activates 12B parameters per token. It joins Nemotron 3 Nano, Nano Omni, Super, Ultra and Nemotron 3.5 Lightning on AWS, deployable via SageMaker Studio or the Python SDK.

AnalysisAI Models1 source

Hugging Face post details HOPE method for pruning MoE experts

A Hugging Face blog post co-published with Amazon introduces higher-order pruning of experts (HOPE) for keeping partner experts together in mixture-of-experts models. The post is titled "Your Experts Have Partners. There's HOPE for Keeping Them Together."

LaunchDevelopers1 source

SageMaker Inference adds Multi-Instance GPU support

AWS says SageMaker GPU endpoints previously required a whole accelerator per model, so an ml.p4d.24xlarge with eight A100s held only eight models. MIG support lets multiple models share a single GPU to raise utilization.