AI Infrastructure · Latest

Amazon Bedrock AgentCore Runtime Enables Live Preview of Coding Agent‑Built Web Apps

Amazon Bedrock AgentCore’s Session Endpoint feature lets a coding agent keep a Vite‑based website running inside an isolated microVM. Each agent session receives an authenticated URL that forwards HTTP and WebSocket traffic to the session’s allowed ports, allowing developers to open the app in a browser, interact with it, and see changes—such as a dark‑theme switch—appear instantly without redeployment. The session endpoint URL is returned in the InvokeAgentRuntime response header X‑Amzn‑Bedrock‑AgentCore‑Runtime‑Session‑Endpoint‑Url, and access requires IAM SigV4 signatures or a JWT bearer token. The sample demonstrates opening a local page with a chat sidebar and the live site, streaming the agent’s steps, and observing internal port handling and request filtering.

Worldwide Express Cuts Freight Exception Costs by 75% with Amazon Bedrock AI

Worldwide Express, a global LTL carrier network with $5 billion in revenue and 130,000 customers, had $15 million of free cash flow tied up each month in unresolved billing exceptions. The company partnered with AWS and LTM to build a multi‑agent AI system on Amazon Bedrock that integrates directly with its SpeedShip transportation management system. The solution automates the three‑step exception workflow—invoice matching, pricing variance, and accessorial validation—by interpreting carrier tariffs and dispute rules. Six months after launch, the system reduced active exceptions by 75 percent, shortened time‑to‑cash by four weeks, saved $1 million annually, and released a one‑time $10 million cash‑flow infusion. The case illustrates how agentic AI can handle complex, context‑dependent freight billing disputes at scale.

Epsilon builds intent-driven UI testing platform with Amazon Nova Act

Epsilon, a data and marketing-services company part of Publicis Groupe, has launched a managed UI testing service using Amazon Nova Act and Amazon Bedrock AgentCore Browser. The platform delivers intent-based testing as a multi-tenant service, allowing product teams to run resilient UI tests via API without maintaining browser infrastructure. A 18-step test workflow demonstrates the capability, which layers intent-based testing on top of Epsilon's existing Playwright regression suite.

MoEngage turns 400 private APIs into agent tools with AgentCore Gateway

MoEngage used Amazon Bedrock AgentCore Gateway to expose roughly 400 internal REST APIs as permission-scoped agent tools without deploying public endpoints. The AI Platform team owns the Model Context Protocol layer, requiring only OpenAPI specs from engineering teams. AgentCore Gateway enforces authorization server-side and provides private connectivity via VPC Lattice, eliminating the need for 20 separate MCP servers and reducing onboarding a new API from days to hours.

NVIDIA RTX Spark laptops gain new models, agent tools and games

NVIDIA RTX Spark laptops, built on Blackwell RTX GPUs and Grace CPUs with up to 128GB unified memory, now include the open-weight speech model Nemotron 3.5 Transcribe for multilingual transcription and translation, the LFM3 family of open models with day-zero support, and GPU resource management controls in the NVIDIA App. New gaming additions include Phantom Blade Zero with ray tracing and DLSS 4.5, and an RTX Remix 1.6 update for Minecraft path tracing.

NVIDIA Nemotron 3.5 Transcribe adds multilingual transcription, speaker attribution and audio Q&A as open model

NVIDIA Nemotron 3.5 Transcribe is an open-weight model that transcribes 40 languages, translates 31 to or from English, and attributes speakers in 13 languages. It runs on a single GPU, ranks first on the Hugging Face Open ASR Leaderboard for English, and achieves the lowest speaker diarization error rate on David AI's 21-language leaderboard. Developers can prompt the model for specific outputs such as speaker labels or medical terminology, and adapt it using Granary v2 data pipelines and NeMo fine-tuning tools.

Property Finder fixes evaluation bias to unlock better models on Amazon Bedrock

Property Finder discovered that 96.8 percent of their evaluation labels were inherited from a single model's predictions rather than verified by human experts, creating systematic bias against alternative models. The team built an unbiased ground truth by manually annotating 356 images and shifted from single-label to multi-label classification. Using evolutionary prompt optimization on Amazon Bedrock, they achieved higher accuracy at 86 percent lower cost.

AWS announces SOCI-indexed images for faster SageMaker AI job starts

AWS introduces Seekable OCI (SOCI) indexing for Amazon SageMaker AI Training and Processing jobs to reduce container image download overhead. With SOCI indexing, SageMaker AI lazy-loads container images, fetching only the files needed to start the container immediately and streaming the rest in the background as code runs. This allows jobs to begin in seconds rather than minutes. Users can adopt the feature by using pre-indexed AWS Deep Learning Containers or by adding an index to their own images, requiring no changes to job definitions. Benchmarks show this reduces GPU idle time during image pulls by up to 90 percent.

Liquid AI releases LFM3 family of specialized multimodal models built on shared 2B parameter backbone

Liquid AI introduces LFM3, a new generation of Liquid Foundation Models ranging from 1.7B to 2B parameters that unify text, vision, audio, omni, and decision models on a single shared backbone. Key improvements include a tri-hybrid architecture combining short convolutions and attention, enabling LFM3-1.7B to outperform the larger LFM2.5-2.6B with 35% fewer parameters. The text model fits within approximately 1 GB at 4-bit precision and supports a 256K context window for on-device agent deployment. Four specialized model variants -- text, vision, audio, and omni -- are derived from the base model and can be fine-tuned on a single GPU. LFM3-1.7B-Base was pre-trained on 22 trillion tokens, with the final 3.5 trillion tokens focused on agentic traces and tool calls to extend the context window.

DeepMind Unveils Genie 4, a Real‑Time World Model with Native Audio Generation

On October 27, 2026 DeepMind announced Genie 4, a multimodal world simulation model that produces photorealistic visuals, fluid physics, and native audio—including music, ambient sounds, and interactive dialogue—simultaneously. Built on the Gemini family, Genie 4 uses Gemini Omni for realistic appearance and behavior, while other Gemini models supply the reasoning needed to keep simulations coherent. Users can steer scenes in real time with natural‑language prompts that alter weather, camera angles, or introduce new objects, with Gemini adapting the context. The model is currently in a research preview for trusted testers, with plans to roll out to Google AI Ultra subscribers and Gemini API partners. Genie 4’s ability to generate realistic physics, human motion, and sound in real time represents a significant step toward safe, infinitely variable simulations for robotics training, AI agents, and new interactive experiences.

C-BRAIN and NVIDIA Use AI to Accelerate Alzheimer’s Research by Unlocking Dark Data

The Consortium for Biomedical Research and AI in Neurodegeneration (C‑BRAIN), a coalition of 18 academic, philanthropic, pharmaceutical and research institutions, has partnered with NVIDIA to build an AI‑driven platform that aggregates previously inaccessible Alzheimer’s data—unpublished findings, negative results and proprietary datasets—into a hybrid centralized/federated system. The platform runs 64 NVIDIA data‑center GPUs on‑premises and additional cloud nodes on Microsoft Azure, with plans to add NVIDIA Blackwell GPUs. By training models on data that never leaves each institution’s secure environment, C‑BRAIN aims to reduce the time for researchers to analyze complex datasets from weeks to hours, thereby speeding the discovery of new treatments for the 57 million people worldwide living with dementia.

Kimi K3.1 Pro launches with 2.8T open weights, narrowing closed AI gap

Moonshot AI launched Kimi K3.1 Pro, a 2.8 trillion parameter open-weights Mixture-of-Experts model with 104 billion active parameters and a 1-million-token context window. Independent benchmarks place the model fourth on the Intelligence Index with a score of 55, surpassing the Fable 5.1 model and ranking eighth in Code Arena. The release includes a Highspeed API offering 2-3x faster output, though at double the cost, and promises open-weight availability on November 4, 2026.

Stream low-latency speech from AI agents using Kokoro TTS on Amazon SageMaker AI

NVIDIA Earth-2 open models deployed across Europe for real-time weather forecasting

NVIDIA Earth-2 provides open AI models for global and regional weather prediction. StormScope EU, a new regional model trained on European weather data, predicts cloud conditions across 16 Meteosat channels every 10 minutes. Developers can customize models using the PhysicsNeMo framework, while organizations like Vaisala Xweather and AXA integrate Earth-2 to enhance weather intelligence and assess cyclone risk.

NVIDIA Cosmos Adds Training, Simulation and Data‑Factory Tools for World Models

NVIDIA has expanded its Cosmos platform with new training utilities, simulation checkpoints and five cloud‑native blueprints in the Cosmos Data Factory. The updates support multimodal, multisensory data curation and synthetic data generation for autonomous vehicles, image augmentation, event video creation, defect imaging and weather‑variant video enhancement. Companies such as Porsche and ANYbotics are already using these blueprints on Azure to cut physical test drives by up to 90% and accelerate development cycles.

NVIDIA Introduces Kumo Relational, a Pre-trained Foundation Model for Enterprise Databases

NVIDIA Kumo Relational is a pre-trained relational foundation model that makes predictions directly on enterprise databases without task-specific training. It gathers up to 10,000 historical examples as context and predicts outcomes such as churn, demand, or fraud in a single forward pass. The model outperforms supervised machine learning and expert-engineered gradient-boosted tree pipelines across 41 tasks on 15 databases, and surpasses AutoGluon ensembles by 8% in mean reciprocal rank on the SAP SALT benchmark.

VocabX method cuts SLM inference cost by 77% through vocabulary specialization

A Portuguese essay grader built on Qwen3-4B reduced average output tokens from 350 to 101 by rebuilding its vocabulary for the specific task. The VocabX method specializes token selection around generation cost rather than frequency, eliminating waste from domain-irrelevant tokens while keeping the base model unchanged.

NVIDIA Unveils RTX Spark, DGX Spark, and DGX Station for Local AI at GTC Berlin

At GTC Berlin, NVIDIA introduced its full lineup of local AI hardware: RTX Spark for Windows PCs, DGX Spark for Linux developers, and DGX Station for enterprise Linux workloads, with a Windows version slated for December. The announcement also revealed new models, software updates, and open‑source workflows, including the Underdog Desktop Agent and the Recursive Flint coding model. These releases aim to enable personal agents, accelerated content creation, and scalable AI development on local machines.

Mistral and NVIDIA partner to accelerate engineering AI with open benchmarks

Mistral and NVIDIA are collaborating to integrate Mistral's AI models with NVIDIA's accelerated computing and physics tools, enabling faster simulations and closed-loop AI-driven design. The partnership will introduce public benchmarks and open datasets to standardize evaluations for physics and agentic AI in computer-aided engineering.

NVIDIA Vera CPUs selected by Nebius, Nscale and Verda to build European AI infrastructure

NVIDIA Vera CPUs will power the next wave of AI infrastructure in Europe. Nebius, Nscale and Verda are adopting the chips to support the computing needs of businesses and developers. Vera accelerates agentic computing tasks such as information retrieval and code execution, helping data centers maintain responsiveness as they scale.