Amazon: latest AI news
All Oossa news about Amazon, newest first.
Known for: Alexa, AWS Bedrock
Follow via RSSToday
Yesterday
Models
Amazon SageMaker adds custom sequence lengths to serverless fine‑tuning
SageMaker AI now lets users set token limits up to 16K when fine‑tuning open‑weight models in serverless jobs.
Models
AWS releases Neuron 2.33 with native PyTorch support
AWS adds TorchNeuron for Trainium, a health‑monitoring Observer, and new debugging tools in a public beta.
Tools & agents
Amazon WorkSpaces adds OAuth‑based AI agent desktop access
AWS now lets AI agents connect to cloud desktops without custom code, using OAuth 2.0 and a managed service.
Business
OpenAI adds Google Cloud to its AI model lineup
OpenAI will let developers run its GPT models on Google Cloud, joining Microsoft and Amazon as hosts.
Tuesday, October 20
Models
Amazon Bedrock Guardrails adds Standard grounding tier
Amazon adds a Standard tier for contextual grounding checks, improving hallucination filtering for conversational AI at no extra cost.
Business
Amazon Bedrock adds optional prompt shortening to Advanced Prompt Optimization
The new setting can cut prompt length by up to 30% while keeping or improving evaluation scores, lowering token costs.
Models
i14 runs Laplace 3B on AWS Trainium2 with 120k tokens/sec training
The 3‑billion‑parameter linear‑attention model trained at 120,000 tokens per second and served 11,544 tokens per second on a trn2.48xlarge instance.
Business
AWS Transform adds support for Control Tower and IaC‑only landing zones
AWS Transform can now work with existing Control Tower setups and generate IaC without direct account access, letting migrations start sooner.
Models
AWS Clean Rooms adds Databricks as a direct data source
AWS Clean Rooms can now query Databricks tables via Delta Sharing, removing the need to copy data to S3.
Models
AWS opens public beta of TorchNeuron for Trainium chips
AWS released TorchNeuron, an open‑source PyTorch backend that runs natively on Trainium AI chips, now in public beta.
Safety
Octave Reliance launches secure multi‑tenant AI foundation on Amazon Bedrock
Octave built a three‑layer AI platform that isolates each customer’s data while embedding 2 billion documents, cutting new‑feature rollout time in half.
Models
AWS adds cross‑Region access to OpenAI GPT‑6 models for Canadian users
Amazon Bedrock now lets developers in Canada call the latest OpenAI GPT‑5.6, GPT‑6 and GPT‑6.1 models via cross‑Region inference.
Models
Amazon Polly adds 14 voices, Hindi and Finnish to Generative engine
The new model, released on Oct 20, 2026, offers more natural speech, low‑latency streaming and expands language support.
Monday, October 19
Models
Amazon Bedrock adds JSON‑schema outputs for Claude models
All Anthropic Claude models on Bedrock can now return schema‑validated JSON, cutting parsing errors and retry logic.
Business
Amazon SageMaker Studio now reachable via simple AWS sign‑up
New developers can open a ready‑to‑use SageMaker Studio notebook in about 20 seconds using Google, GitHub, Apple or Amazon accounts, and get $100 free credits.
Business
Grok 4.7 added to Amazon Bedrock in AWS GovCloud (US)
Amazon Bedrock now offers SpaceXAI's Grok 4.7 model for coding, agentic tasks, and knowledge work in the US GovCloud region.
Business
Amazon CloudWatch Omni expands to eight new AWS regions
AWS adds CloudWatch Omni, its AI‑driven observability tool, to Canada, three Asia Pacific locations, two European sites and São Paulo.
Business
EC2 Image Builder can now use Auto Scaling Group instances
AWS adds an optional feature that lets Image Builder pick already‑running EC2 instances from an Auto Scaling Group, cutting build time.
Business
Amazon GuardDuty adds optional runtime detection rules
GuardDuty now offers over 50 optional rules for Runtime Monitoring on EC2, EKS and ECS, with four new finding types.
Models
Push‑based KV transfer cuts first‑token latency in vLLM
AWS shows the new push mode can make first token appear up to three times faster and reduce per‑token time by 30% for the Qwen3‑32B model.
Models
Deepgram cuts speech‑to‑text costs 11‑70% with AWS Trainium2
By porting its Nova‑3 model to AWS’s Trainium2 chips, Deepgram processes more audio per dollar on batch jobs.
Models
AWS cuts first‑token latency for multi‑turn AI agents with bi‑directional KV transfer
Bi‑directional key‑value cache sharing on EC2 P5en GPUs lowers TTFT up to 3.3× for long conversations.
Business
TorchNeuron adds native PyTorch support for AWS Trainium
PyTorch releases TorchNeuron, an open‑source backend that lets models run on Trainium chips with unchanged code.
Friday, October 16
Business
Replit adds Bring Your Own Cloud options for enterprises
Replit now lets companies run the full platform or just published apps in their own Azure, Google Cloud, or AWS tenant.
Business
AWS Transform for migrations launches in Singapore region
AWS now offers its AI‑driven migration service in Asia Pacific (Singapore), giving customers local data residency and lower latency.
Business
AWS Lambda durable functions launch in GovCloud US regions
AWS adds durable functions to its GovCloud (US‑West and US‑East) regions, letting government users build multi‑step apps without extra compute charges.
Models
NVIDIA Nemotron 3.5 Super VL comes to SageMaker JumpStart
AWS customers can deploy NVIDIA’s model for image, video and text tasks, with one variant for inference and another for customization.
Tools & agents
Brands face surge of AI shopping agents before holidays
Cheq reports a 100‑fold rise in AI agent traffic as Instinct and Meta’s Muse launch, while Amazon and eBay block them.
Models
Reflection’s Beam model comes to AWS SageMaker and Trainium3
AWS says businesses can deploy Reflection’s Apache 2.0-licensed open-weight model in their own AWS accounts, initially in two US regions.