AWS AI Updates: September 26, 2026

1. DeepEP Over EFA Lifts MoE Reinforcement Learning Rollout Throughput by 40 Percent

AWS. AWS described how it scaled RL post-training (RLHF and GRPO) for a super-sparse Mixture-of-Experts model on Amazon EKS across 48 P5en instances, 16 for training and 32 for inference. Replacing generic all-to-all collectives with DeepEP’s dispatch and combine kernels, now running natively on EFA through an Amazon-contributed libfabric migration, raised aggregate rollout throughput by 40 percent over a Slime-based baseline without EFA-accelerated expert parallelism. The stack pins CUDA 13.0, PyTorch 2.12.1, NCCL 2.31.2, DeepEP 2.0.0, and SGLang 0.5.17, and communicating nodes must sit in the same Availability Zone. Source

2. SkyRL on HyperPod Takes a Qwen3-VL-8B Maze Solver From 44 to 97 Percent With Six GPUs

AWS. A walkthrough runs the open-source SkyRL framework on SageMaker HyperPod to apply GRPO to Qwen3-VL-8B for multi-turn visual maze navigation, starting from a VisGym SFT checkpoint. The cluster uses three ml.g7e.12xlarge workers (six RTX PRO 6000 Blackwell GPUs total) with the policy sharded via FSDP and six colocated vLLM engines generating rollouts. Solve rate on a fixed 64-maze held-out set rose from 43.75 percent to 96.875 percent by step 160, with 8 samples per prompt and a 15-turn cap. Source

3. AWS’s Internal Executive Assistant Validates Numbers While the LLM Is Still Streaming

AWS. NarrateAI, an internal assistant on Bedrock AgentCore serving more than 4,000 AWS executives, evaluates each paragraph in a producer-consumer pipeline during generation instead of after it, cutting latency 86.8 percent versus sequential post-generation checks. Data accuracy runs as a cascade, with exact value matching at 0.3 ms and semantic verification at about 1,758 ms reserved for mismatches, which cut verification cost 54 percent. AWS reports 99.3 percent numerical accuracy over six months in production, plus cross-account multi-model failover to spread load across quota pools. Source

4. Qwen3-TTS Voice Cloning Fits on a Single L4 GPU on SageMaker AI

AWS. AWS published a deployment of Qwen3-TTS-12Hz-1.7B-Base on an ml.g6.4xlarge (one 24 GB NVIDIA L4) using the vLLM-Omni container, with the talker and code2wav stages sharing one GPU by setting SM_VLLM_GPU_MEMORY_UTILIZATION to 0.45. The model clones a voice from a few seconds of reference audio plus transcript and supports cross-lingual cloning, emitting 24 kHz mono WAV. The post gives no latency or throughput figures, and requests must carry the route=/v1/audio/speech attribute or they are rejected. Source

5. Qumulo Caching Lets HyperPod Train in One Region on Data Stored in Another

AWS. A reference setup keeps the training dataset in a Qumulo hub in us-east-2 and trains on a HyperPod cluster in us-west-2, with Qumulo’s NeuralCache prefetching to local NVMe across 60 ms of inter-region latency. Training a 1.02B-parameter Llama 3 model on two ml.p5.48xlarge instances, the warm cross-region spoke hit 115 to 116 samples per second against 116 to 117 in-region, with a 94 to 96 percent cache hit rate. The first 100 to 150 batches run 15 to 20 percent slower while the cache warms. Source

6. SES and End User Messaging Ship Agent Skills for the AWS MCP Server

AWS. Amazon SES and AWS End User Messaging now publish AI agent skills for the AWS MCP Server that walk a coding agent through tasks such as verifying sending identities, sending production email, and building branded RCS agents. The skills work with Claude Code, Codex, Cursor, and Kiro; in the first three, the aws-core plugin bundles the MCP server and a curated skill set in one install. Source

7. Transcribe Custom Language Models Can Now Be Encrypted With Customer-Managed KMS Keys

AWS. Amazon Transcribe now accepts a customer-managed symmetric KMS key when creating or updating custom vocabularies, vocabulary filters, and custom language models, replacing the previously mandatory AWS-owned key. Key use is logged in CloudTrail, and disabling the key revokes access to the stored artifacts. The change is opt-in and available in all Regions where Transcribe runs. Source