AWS AI Updates: August 11, 2026

1. FLUX.2 Decoder and Gemma 4 Multimodal Model Land on SageMaker JumpStart

AWS. Amazon SageMaker JumpStart added Black Forest Labs’ FLUX.2-small-decoder and Google’s gemma-4-12B-it to its model catalog. FLUX.2-small-decoder is a distilled VAE decoder that acts as a drop-in replacement for the standard FLUX.2 decoder, delivering roughly 1.4x faster decoding at 1.4x lower VRAM consumption and helping teams cut cost on high-resolution image generation. The gemma-4-12B-it model uses an encoder-free architecture that routes text, image, and audio inputs into a single decoder-only transformer, runs on 16GB of RAM, and natively supports function calling, giving practitioners a memory-efficient multimodal option for agentic workflows. Source

2. Redis, JetBrains, and LightOn Models Target Caching, Coding, and OCR

AWS. SageMaker JumpStart now offers three specialized foundation models aimed at distinct production workloads. Redis’s langcache-embed-v3-small is tuned for semantic caching in LLM applications, converting text to dense vectors to catch semantically similar queries and cut redundant model calls. JetBrains’ Mellum2-12B-A2.5B-Thinking is a Mixture-of-Experts coding model that activates only 8 of 64 experts per token, using 2.5B of its 12B parameters with a 131,072-token context and chain-of-thought reasoning, while LightOn’s 1B-parameter LightOnOCR-2-1B handles end-to-end multilingual document-to-text conversion without a traditional OCR pipeline. All three deploy from the JumpStart catalog or the SageMaker Python SDK. Source

3. GLM-5.2, Nemotron Nano, and GLM-OCR Add Long-Context and Reasoning Options

AWS. Amazon added GLM-5.2 FP8, NVIDIA-Nemotron-Nano-12B-v2, and GLM-OCR to SageMaker JumpStart, broadening the catalog for long-horizon and document workloads. GLM-5.2 FP8 targets agentic engineering tasks with a usable 1M-token context window spanning requirements through deployment, while Nemotron-Nano-12B-v2 pairs a hybrid Mamba-2 and Transformer design with a 128K context and claims up to 6x higher inference throughput than comparable models. GLM-OCR, at 0.9B parameters, reconstructs scanned PDFs, handwritten notes, and multilingual documents into Markdown, JSON, or LaTeX, and each model deploys in a few clicks from the JumpStart catalog or via the SageMaker Python SDK. Source