AWS Adds Eight Open Models to SageMaker JumpStart

Amazon has expanded SageMaker JumpStart with a batch of eight third-party open models spanning image generation, OCR, coding, embeddings, and general-purpose reasoning, giving developers pre-configured access to deploy them directly within AWS infrastructure.

The lineup covers a wide range of use cases. FLUX.2-small-decoder targets faster image generation pipelines, claiming roughly 1.4x quicker decoding at 1.4x lower VRAM usage compared to prior variants. Gemma-4-12B-it, meanwhile, is pitched as a compact multimodal model capable of running on just 16GB of RAM while handling text, image, and audio inputs alongside native function calling for agentic workflows.

On the document-processing side, GLM-OCR and LightOnOCR-2-1B stand out for their small footprints. GLM-OCR, at 0.9B parameters, converts documents including tables and formulas into Markdown, JSON, or LaTeX. LightOnOCR-2-1B is roughly 9x smaller than competing OCR approaches yet reportedly achieves state-of-the-art results on the OlmOCR-Bench benchmark.

Two models are aimed squarely at long-context and engineering-heavy workloads. GLM-5.2 FP8 offers a 1-million-token context window designed for project-scale agentic engineering tasks, while JetBrains' Mellum2-12B-A2.5B-Thinking uses a Mixture-of-Experts design, activating only 2.5B of its 12B total parameters per forward pass across 64 experts (8 active at a time), with a 131,072-token context length.

NVIDIA-Nemotron-Nano-12B-v2 rounds out the reasoning-focused additions with a hybrid Mamba-2 and Transformer architecture, claiming up to 6x higher inference throughput than comparable open models at a 128K context length. Redis contributed langcache-embed-v3-small, an embedding model built for semantic caching in LLM application pipelines.

Naming Discrepancies Flagged

Independent verification of two entries in AWS's announcement ran into inconsistencies. Public records point to a "gemma-4-E2B-it" variant appearing on JumpStart in July 2026, not the "gemma-4-12B-it" named in AWS's January posting, and no external source corroborates a distinct 12B version under that exact name.

Similarly, outside reporting identifies a "FLUX.2-klein-base-4B" model added to JumpStart in May 2026, but no independent source confirms the "FLUX.2-small-decoder" variant AWS describes. The specifications may reflect legitimate model updates or renamed releases, but the exact variant names in AWS's announcement remain unverified against third-party sources.

For teams building on SageMaker, the practical takeaway is straightforward: JumpStart continues to broaden its catalog of smaller, efficiency-oriented open models suited to constrained hardware and specialized tasks like OCR and semantic caching, though buyers should confirm exact model specifications before deployment given the naming ambiguities.