AWS AI Updates: July 14, 2026

1. OpenAI GPT-5.6 Sol, Terra, and Luna Reach General Availability on Amazon Bedrock

AWS. Three OpenAI GPT-5.6 variants are now generally available on Amazon Bedrock: Sol, a flagship reasoning model that AWS says delivers state-of-the-art results on agentic coding benchmarks; Terra, a balanced tier positioned at GPT-5.5-level performance for about half the cost; and Luna, a cost-efficient model for fast, low-priced inference. The models support prompt caching with explicit cache breakpoints, billing repeated context across agentic workflows at a 90 percent discount, which matters for long-running coding agents and other high-context loops. Access runs through AWS infrastructure and IAM, and usage counts toward existing AWS commitments. Source

2. Gemma-4-E2B-it Lands in Amazon SageMaker JumpStart

Amazon SageMaker. Google DeepMind’s Gemma-4-E2B-it, a multimodal instruction-tuned model, is now available in SageMaker JumpStart. It accepts text, image, and audio input and generates text, with a built-in reasoning mode that lets it work through problems step by step before answering, plus strengths in object detection, document analysis, UI understanding, OCR, video analysis, and function calling for agent workflows. The model offers multilingual support across dozens of languages and can be deployed from SageMaker Studio or the SageMaker Python SDK. Source

3. Qwen3 Embedding and Reranking Models Arrive in SageMaker JumpStart

Amazon SageMaker. Two Qwen3 retrieval models are now in SageMaker JumpStart for building search and RAG pipelines. Qwen3-VL-Embedding-2B generates vectors from text, images, screenshots, videos, or mixed-modality inputs and supports image-text retrieval, video-text matching, and visual question answering across more than 30 languages. Qwen3-Reranker-4B takes a query and document pair and returns a relevance score for text and code retrieval, classification, and clustering across more than 100 languages, accepting user instructions for task-specific tuning. AWS positions the pair to work together, with the embedding model handling initial recall and the reranker refining results. Both deploy from SageMaker Studio or the Python SDK. Source

4. Voxtral-Mini-4B-Realtime Brings Streaming Speech-to-Text to SageMaker JumpStart

Amazon SageMaker. Mistral AI’s Voxtral-Mini-4B-Realtime is now available in SageMaker JumpStart for real-time speech transcription. A natively streaming architecture enables low-latency transcription across 13 languages, and configurable transcription delays let developers trade off latency against accuracy for their use case. The model can be deployed from SageMaker Studio or the Python SDK, targeting applications that need live speech processing such as captioning, voice assistants, and meeting transcription. Source

5. OpenAI privacy-filter Adds PII Detection and Masking in SageMaker JumpStart

Amazon SageMaker. OpenAI’s privacy-filter, a bidirectional token-classification model for detecting and masking personally identifiable information in text, is now available in SageMaker JumpStart. It labels an input sequence in a single forward pass to identify categories including account numbers, addresses, emails, names, phone numbers, URLs, dates, and secrets, and AWS describes it as fast, context-aware, and tunable for high-throughput data sanitization. Teams can deploy it from SageMaker Studio or the Python SDK to build PII redaction into data pipelines and LLM preprocessing. Source

6. SageMaker HyperPod Adds Custom AMIs for Slurm Clusters

Amazon SageMaker. HyperPod now lets teams deploy Slurm clusters from custom Amazon Machine Images, so security teams can bake compliance tools, security agents, proprietary libraries, and specialized drivers directly into pre-approved base images. Custom AMIs are specified through the CreateCluster, UpdateCluster, or UpdateClusterSoftware APIs, and AWS says the approach replaces complex lifecycle configuration scripts that slow deployment and create inconsistencies across nodes, yielding faster startup and more reliable, consistent environments. The feature keeps compatibility with distributed training libraries and HyperPod cluster management. Source