AWS AI Updates: October 22, 2026

1. Reflection’s Open-Weight Beam Model Arrives on SageMaker JumpStart, Optimized for Trainium3

AWS. Beam, the first open-weight model from US-based Reflection, is now available in Amazon SageMaker JumpStart under the Apache 2.0 license, and AWS says it was ported and optimized for Trainium3 from day one. The model ships in BF16, FP8, and NVFP4 checkpoint variants and deploys to a fully managed JumpStart endpoint, initially in US East (N. Virginia) and US East (Ohio). Reflection positions Beam for reasoning, coding, and agentic work, and the open weights let teams fine-tune, inspect, and modify the model inside their own AWS accounts rather than calling a hosted API. Source

2. Neuron 2.33 Adds Neuron Observer Health Detection, vLLM 0.29, and 25 New NKI Kernels

AWS. The AWS Neuron 2.33 SDK release packages the TorchNeuron PyTorch backend for Trainium (public beta) and introduces Neuron Observer, also in public beta, which detects Trainium health issues in under a second and identifies the affected node or device. The vLLM Neuron plugin moves to vLLM 0.29.0, a vLLM Omni plugin for Wan2.2 video generation joins the release, and new Deep Learning AMI and container images bundle PyTorch 2.14 and TorchTitan. Neuron Explorer can now open torch.profiler captures and attribute memory allocations to Python source lines for debugging out-of-memory errors, while the NKI Library adds 25 experimental kernels, including ones for DeepSeek-V3.2, Qwen3.5, and Kimi Delta Attention. Source

3. SageMaker Serverless Model Customization Lets You Set Sequence Lengths

AWS. SageMaker AI serverless model customization jobs now accept configurable sequence lengths instead of the fixed values previously baked into each fine-tuning recipe. Users set dataset_max_len for SFT and DPO jobs, or max_prompt_length and max_response_length for RFT jobs, up to a published per-model maximum of as much as 16K tokens for many models. The feature covers open-weight families such as gpt-oss, Gemma, Llama, Nemotron, and Qwen, and is available in US East (N. Virginia), US West (Oregon), Asia Pacific (Tokyo), and Europe (Ireland). Source

4. WorkSpaces for AI Agents Moves Authentication and Session Handling Into Its Managed MCP Service

AWS. Amazon WorkSpaces for AI agents now handles authentication and desktop connections inside its managed Model Context Protocol service, so agents create sessions with existing AWS credentials over OAuth 2.0 instead of custom code that generated and rotated streaming URLs. Operators can query which stack, fleet, and session an agent is attached to and generate a link to watch the agent in real time. A new catalog MCP tool lets agents list and launch desktop applications programmatically rather than navigating menus, which AWS says improves accuracy and reduces token cost. Source