AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Amazon SageMaker JumpStart: 2026 year-to-date in review

Collected Oct 19, 2026

Amazon Web Services published a year-to-date review of Amazon SageMaker JumpStart covering January through September 2026. JumpStart is the model catalog inside SageMaker AI that bundles model cards, solution templates and example notebooks, and the review reports that more than 100 new models from 30 distinct providers were added during that window. Alongside the catalog growth, AWS shipped three console and workflow changes aimed at the path from picking a model to running it: a Model Hub surfaced on the main SageMaker AI console page, named deployment configurations with optimization profiles, and a Hugging Face integration that deep-links into SageMaker Studio.

JumpStart is organized around three jobs: discovery and evaluation of foundation models, one-click deployment to a SageMaker real-time inference endpoint with pre-configured settings, and customization. Fine-tuning runs from the console or the SageMaker Python SDK and supports LoRA, supervised fine-tuning, SFT, DPO and RLVR, with serverless options that remove much of the infrastructure management. Model cards carry architecture details, benchmarks, licensing and sample notebooks.

The provider breakdown is the clearest signal of where the catalog is weighted. Alibaba's Qwen contributes 29 models, the largest single block, spanning reasoning (Qwen 3.8-27B), a large mixture-of-experts model (Qwen3.5-122B-A10B), agentic workflows (Qwen-AgentWorld-35B-A3B), code generation (Qwen3-Coder-Next, Qwen 3 Coder 30B), speech recognition (Qwen3-ASR-1.7B), text-to-speech (Qwen3-TTS), embeddings (Qwen3-VL-Embedding-2B) and reranking (Qwen3-Reranker-4B). NVIDIA adds 17, including the Cosmos3 world foundation model family in Super, Edge and Nano tiers, Cosmos Reason 7B, MSA-Search, Nemotron 3.5 Lightning, Nemotron-3-Super-120B in both FP8 and BF16, Nemotron3 Ultra, Nemotron-Nano-12B-v2, the multimodal Nemotron 3 Nano Omni and Nemotron 3.5 Super VL. Mistral supplies 7, among them the Ministral-3 family at 3B, 8B and 14B in standard and BF16 forms plus Voxtral-Mini-4B-Realtime for speech. Google contributes 6 Gemma 4 variants (gemma-4-31B-it, E2B-it, E4B-it, 26B-A4B-it, an assistant build and an NVFP4 quantized version). IBM adds 4, including Granite 4.0 h-micro, h-small and h-tiny and granite-speech-4.1-2b for multilingual speech recognition across six languages. Cohere adds 4, including Command A (A100) and Rerank v3.5, v4.0 Fast and v4.0 Pro. A further 33 models come from MiniMax, Zhipu AI (GLM), DeepSeek, Moonshot AI, Allen AI, Microsoft, Meta (Muse Spark), Reflection (Beam), Black Forest Labs, JetBrains, OpenBMB, BharatGenAI, Kakao, OpenFold and others.

AWS groups the additions into a handful of themes. NVIDIA's contribution is framed as reaching beyond language: Cosmos3 is described as combining video generation, physics simulation and action planning in one architecture for physical AI such as robotics, autonomous systems and digital twins, while the Nemotron line runs from edge-sized inference at Nano-12B to large-scale reasoning at Super-120B.

Multimodality is a second theme. Nemotron 3 Nano Omni handles text, vision and speech in a single model; Gemma 4 adds vision-language capability across sizes; granite-speech-4.1-2b covers recognition and translation in six languages; Voxtral-Mini supports real-time speech understanding. The practical effect AWS cites is that customers can stop chaining separate preprocessing models for multimodal work, which in a typical pipeline means fewer hops, fewer failure points and less glue code between services.

Creative generation is treated as a mainstream category now: Black Forest Labs' FLUX.2 Small Decoder and FLUX.2-klein-base-4B for fast text-to-image, Meta Muse-Glimmer-30B for multimodal agentic workflows, and Cosmos3 for video. Efficiency gets its own group: Zhipu AI GLM-5.2 FP8 for a smaller memory footprint, JetBrains Mellum2-12B with a mixture-of-experts design of only 2.5B active parameters for code generation, OpenBMB MiniCPM5-1B packing multimodality into one billion parameters, and Microsoft Phi-4-mini-instruct and IBM Granite 4.0 h-tiny positioned as production-capable on instances like ml.g5.xlarge. This is a sensible push, since the dominant cost in serving is often the instance rather than the model license, and smaller weights let teams fit a model on a cheaper GPU class.

Purpose-built models form another cluster: Qwen-AgentWorld-35B-A3B for autonomous workflows with native tool-calling; GLM-OCR, LightOnOCR and DeepSeek-OCR for document text extraction; Qwen3-Reranker-4B and Cohere Rerank v4.0 for retrieval precision in RAG pipelines; langcache-embed-v3-small for lightweight semantic search embeddings; Qwen3-ASR and Qwen3-TTS for speech in and out; and OpenFold3 for protein structure prediction. The implied architecture is a pipeline of specialists rather than one general model doing everything, which is usually cheaper to tune but adds orchestration overhead.

On the tooling side, the Model Hub now appears on the SageMaker AI console landing page instead of being buried inside SageMaker Studio's left navigation. Browsing supports filters for modality, publisher, model size, license and framework, and notebooks or models open in Studio with one click. Deployment gains four named profiles: Balanced for a general cost, latency and throughput tradeoff; Throughput optimized for maximum tokens per second using techniques such as Speculative Decoding, aimed at high-volume batch work; Latency optimized for the lowest time to first token in interactive applications; and Cost optimized for the lowest infrastructure cost per request where higher latency is acceptable. Those profiles include Neuron-optimized setups for AWS Inferentia and AWS Trainium instances. The pitch is skipping weeks of manual benchmarking across instance types, containers and settings.

The Hugging Face integration links catalog entries on the Hugging Face Model Hub to a configured Studio environment. Selecting Deploy on SageMaker AI or Customize and Deploy on SageMaker AI lands the user in Studio with the model loaded, permissions set and GPU quota visible, avoiding manual IAM role, domain and permission setup before experimenting. Access works through the Studio console (Models, then JumpStart models), the SageMaker Python SDK, or the Hugging Face deep link.

A likely trade-off is version drift: a fast-moving catalog of 100-plus models in nine months means teams need to pin specific model builds and re-test before swapping, since benchmarks on a model card rarely match a production workload. The deployment profiles also encode assumptions about traffic shape that may not hold once request mix changes.

Why it matters: Teams building on AWS get a materially broader menu of open models, including compact and quantized options that fit cheaper GPU instances, plus defaults that reduce the setup work between choosing a model and serving it. In practice, the Hugging Face deep link is likely the most consequential change for individual developers, since it removes the account plumbing that often precedes a first experiment; that is an inference, not a stated result. Enterprise buyers weighing vendor concentration also gain more named providers to choose from, though the heavy Qwen weighting means the catalog's center of gravity still leans toward one family.

Read at AWS Machine Learning Blog

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Amazon SageMaker JumpStart onboarded over 100 new models from 30 providers in the first three quarters of 2026. This recap covers the model-onboarding highlights and JumpStart enhancements, including the simplified Model Hub, pre-configured deployment options, and Hugging Face deep link integration.