AWS AI Updates: May 15, 2026

1. Amazon Bedrock launches advanced prompt optimization across multiple models

AWS. Bedrock added a prompt optimization tool that takes a template, sample inputs, optional ground-truth answers, and evaluation criteria, then iteratively refines the prompt across up to five foundation models in parallel, including multimodal inputs (JPG, PNG, PDF). Each run returns evaluation scores, cost estimates, and latency measurements, replacing the days-to-weeks manual loop teams typically run when migrating models or chasing better quality. Source

2. SageMaker AI adds serverless fine-tuning for Qwen3.6 27B

AWS. SageMaker now offers serverless customization of Alibaba’s Qwen3.6 27B with both supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT), with AWS handling infrastructure provisioning and training orchestration while customers pay only for usage. Access is via SageMaker Studio or the Python SDK in N. Virginia, Oregon, Tokyo, and Ireland, removing cluster management as a prerequisite for domain adaptation. Source

3. SageMaker JumpStart adds FLUX.2-klein image model and Qwen3 multilingual embeddings

AWS. JumpStart now offers Black Forest Labs’ FLUX.2-klein-base-4B, a compact image generator that runs on as little as 13 GB of VRAM and supports multi-reference editing, plus Qwen3-Embedding-0.6B, an instruction-aware embedding model covering 100+ languages. The pairing targets creative content pipelines and multilingual RAG/semantic search, with one-click or SDK deployment inside SageMaker Studio. Source

4. SageMaker JumpStart adds Qwen3 speech models for TTS and ASR

AWS. Three new Qwen3 speech models landed in JumpStart: TTS-12Hz-1.7B-CustomVoice (10-language TTS with tone, feeling, and speech-pattern controls), TTS-12Hz-1.7B-Base (multilingual TTS with 3-second voice cloning), and ASR-1.7B (52-language automatic speech recognition tuned for challenging audio). The set is deployable in a few clicks, giving teams a single source for voice assistants, captioning, transcription, and multilingual support without stitching together separate vendors. Source

5. SageMaker JumpStart adds GLM-5.1-FP8 for agentic coding and Phi-4-mini for edge reasoning

AWS. Z.ai’s GLM-5.1-FP8 arrived in JumpStart with a focus on repository-level code generation, terminal tasks, and complex debugging, supporting extended reasoning across hundreds of iterations for agentic dev workflows. It ships alongside Microsoft’s Phi-4-mini-instruct, a compact 24-language model with function calling aimed at memory-constrained, latency-bound deployments, both deployable through JumpStart with minimal configuration. Source