AWS AI Updates: October 17, 2026

1. Reflection’s Beam Open-Weight Coding Model Launches on SageMaker JumpStart and Trainium3

AWS. Beam, the first open-weight model from US-based Reflection, is now available in SageMaker JumpStart in US East (N. Virginia) and US East (Ohio). It is a sparse mixture-of-experts model with 501B total and 23B active parameters, a reported 1M-token effective context, and Apache 2.0 weights on Hugging Face in BF16, FP8, and NVFP4 variants, aimed at resolving software issues, tool calling, and long multi-step agent workflows. AWS says the model was ported and optimized for Trainium3 from day one, and its deployment guide covers JumpStart real-time endpoints plus SageMaker HyperPod Inference with SGLang or vLLM (FP8 on ml.p5/p5e/p5en H100 and H200 instances, NVFP4 on ml.p6-b200.48xlarge). Source

2. NVIDIA Nemotron 3.5 Super VL Arrives on SageMaker JumpStart in NVFP4 and BF16 Variants

AWS. NVIDIA’s Nemotron 3.5 Super VL vision-language model is now deployable from SageMaker JumpStart. It pairs a 120B-parameter hybrid MoE backbone (12B active) with a vision encoder, accepts image, video, and text input with up to 1M tokens of context, and includes a Multi-Token Prediction head for speculative decoding plus per-request configurable reasoning, tool calling, and structured JSON output. JumpStart offers an NVFP4 variant for inference on G7 instances with RTX PRO 4500 Blackwell Server Edition GPUs and BF16 reference weights on ml.p5en.48xlarge for SFT, RL, and distillation work. Source

3. AWS Transform’s .NET Agent Now Converts Azure SDK Code to AWS Services

AWS. AWS Transform made application code conversion from the Azure SDK to the AWS SDK generally available, extending its .NET modernization agent to Azure cloud-native apps. The agent detects Azure service calls in submitted source, proposes mappings in its assessment report and modernization plan for the user to accept or adjust, then rewrites and validates the code: Azure Functions to Lambda, Blob Storage to S3, and Queue Storage to SQS are GA, with Cosmos DB to DynamoDB in preview. The feature is available in the eight Regions where AWS Transform runs, including US East (N. Virginia), Frankfurt, London, and Tokyo. Source

4. AWS Details the Nine-Stage AgentCore Evaluations Pipeline Behind Its Internal Sales Agents

AWS. AWS’s Sales, Marketing, and Global Services Operations team published the evaluation framework it runs on Bedrock AgentCore Evaluations for a multi-agent system serving AWS field sales. It combines AgentCore built-in evaluators, six custom failure-type evaluators (metric swaps, wrong period, wrong customer, hallucination, and more), DeepEval and AutoEval, and deterministic ground-truth checks. A quality gate scores 200 held-out questions with an LLM-as-judge, and Online Evaluation samples 10 percent of live traffic for drift detection. The team reports pure hallucinations falling from 138 to under 52 per 10,000 responses after using AgentCore optimization’s system prompt recommendations, with regressions now detected in under four hours instead of days or weeks. Source

5. AWS Open-Sources Strands Decider 2B and Previews the Strands Box Agent Sandbox

AWS. AWS released two open-source projects for agent builders. Strands Decider 2B is a 2-billion-parameter model built on Qwen3.5-2B with its language-model head replaced by a pointer head, so it picks among predefined options (model routing, tool selection, classification, guardrails) and returns confidence scores instead of generating text; AWS published the weights, training data, and scripts, and reports a median decision latency of about 115 ms on an RTX 3090 and 72.3% accuracy on JevBench’s 231 public tasks. Strands Box, in developer preview, is a kernel-isolated sandbox for shell, Python, MCP, filesystem, and network actions, governed by the open-source Dogwood policy language, which can enforce rules such as “tests must pass before pushing code” outside the model and works with other agent frameworks. Source