AI news for builders and product teamsUpdated Nov 17, 2026, 13:01 UTC
AWS Machine Learning Blog
First-party releases and research from AWS Machine Learning Blog. Headlines and excerpts link to the original articles.
Latest stories
Newest first
AWS says Amazon Connect has been named a Leader in Gartner's 2026 Magic Quadrant for Contact Center as a Service, the fourth consecutive year. The Gartner document, attributed to Drew Kraus and others, is dated 19 October 2026.

AWS added Seekable OCI (SOCI) indexing to SageMaker AI Training and processing jobs, so container images lazy-load and jobs start in seconds. In AWS benchmarks, image pull times fell from minutes to 3–11 seconds, cutting GPU idle time during pulls by up to 90 percent.

Property Finder found that 96.8 percent of its vision-model evaluation labels were copied from the incumbent model's own predictions, biasing every comparison; after rebuilding ground truth with expert annotations, it switched to Meta's Llama 4 Maverick 17B Instruct on Amazon Bedrock, reaching Micro F1 0.80 versus 0.56 for its prior model at an estimated 86 percent lower cost.

Exowatt fine-tuned Qwen3 models on SageMaker serverless training and beat frontier models on two energy tasks, scoring 0.911 macro-F1 for edge alarm triage versus 0.698 for Claude Sonnet 4.6 and 0.975 for cloud market analysis versus 0.625 for Claude Opus 4.7, the latter at about 49x lower cost per answer.

AWS's Machine Learning Blog documents prompt caching patterns for OpenAI models on Amazon Bedrock, reporting that implicit caching cut input cost about 62 percent versus no caching on GPT-6 Astra in a six-turn test workload and explicit rolling breakpoints saved about 61 percent.

Bayer Crop Science built the Agentic Decision Engine, a 22-agent natural-language layer over its ML workflow platform, running on Amazon Bedrock AgentCore; it resolves about 80 percent of user questions without an engineer and has been used by more than 100 people since launch.

AWS has upgraded the Amazon Polly Generative engine with 14 new voices across 9 locales, including the first generally available generative Hindi and Finnish voices, now reaching 57 voices in 11 languages and 25 locales. Existing integrations get the improved audio with no code or API changes and no price increase.

IskayPet, a pet care company with more than 330 stores across Spain, Portugal and France, built a multilingual corporate knowledge assistant on Amazon Bedrock Managed Knowledge Bases and reports that time to find an operational answer fell from 3–5 minutes to about 7 seconds in its first two months in production.

AWS made training sessions for Amazon SageMaker AI serverless model customization generally available, letting teams write their own post-training loop — custom loss functions and reward signals — while SageMaker handles GPU allocation, distributed execution, fault recovery and checkpoint storage. It supports four open-weight models: gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B and Qwen3.8-27B.

Amazon Quick is now a HIPAA-eligible service across its web, desktop and mobile clients, letting healthcare organizations process protected health information under a signed AWS Business Associate Addendum — provided they configure encryption, access controls and audit logging correctly.

AWS published a walkthrough for streaming each sentence from an agent to a hosted Kokoro-82M text-to-speech model on a SageMaker AI real-time endpoint as soon as the sentence exists, cutting mean time to first audio byte from 329.3 ms to 156.7 ms in isolation.

An AWS sample project replaces guessed SageMaker AI auto scaling targets for text-to-speech with values derived from two live measurements: how many concurrent requests one instance serves within a 3,000 ms latency SLO, and how long a real scale-out takes.

Rescale built an AI triage agent for nightly Playwright test failures using AWS's Strands SDK and the AI-DLC methodology, going from problem definition to staging in five days; classification agreement with human triage rose from 44% to 68%.

AWS's Solutions Architecture team built Deep Insight, an open-source multi-agent data analysis system, and documented the five production challenges it addresses in the first post of a three-part series. It combines the Strands Agents SDK, Amazon Bedrock models per agent role, and Bedrock AgentCore for deployment, with sandboxed code execution on Fargate.

AWS's Adaptive Agentic Intelligence Framework, built on Strands Agents, uses a 1B-parameter llama3.2 orchestrator running locally via Ollama to route routine requests to on-device tools and complex ones to Amazon Bedrock models.

AWS showed how to train Qwen3 three ways on a single trn2.48xlarge, using Transformers v5 as a modeling backend, torchtitan's native Qwen3 model, or NKI mixture-of-experts kernels with custom overrides.

U.S. Hunger has given 17 employees access to Amazon Quick, using daily 2 AM ET SPICE refreshes from SQL Server and shared Spaces to let its CEO and leadership team explore program and financial data without filing new data-pull requests.

i14 and AWS trained and served the Laplace 3B linear attention model on a single Trainium2 instance using native PyTorch via TorchNeuron Native, reaching about 120,000 training tokens per second and 11,544 output tokens per second across 1,280 concurrent sequences.

Smartsheet has released a generally available Model Context Protocol server that lets AI assistants read, write and manage Smartsheet data through more than 40 tools, with setup paths for Amazon Quick and the Kiro IDE.

AWS published an end-to-end pattern for conversational supply chain intelligence that uses Amazon Bedrock AgentCore as the tool and orchestration layer behind Amazon Quick, with answers pulled from live systems of record instead of model recall. An open-source sample walks through four business MCP servers, a discovery registry, and human-approved actions.