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

Amazon SageMaker JumpStart is a machine studying (ML) hub inside Amazon SageMaker AI that gives entry to a curated catalog of AI fashions, answer templates, and instance notebooks. It reduces the time it takes to go from mannequin choice to deployment. You need to use it to guage open supply giant language fashions (LLMs), deploy a imaginative and prescient mannequin for doc understanding, or fine-tune a compact mannequin by yourself information.

SageMaker JumpStart serves three main use instances:

Mannequin discovery and analysis — Browse a whole lot of basis fashions (FMs) from suppliers like Alibaba (Qwen), NVIDIA, Google, Mistral, Cohere, IBM, and extra. Every mannequin card consists of structure particulars, benchmarks, licensing info, and pattern notebooks.
One-click deployment — Deploy your alternative of mannequin to a SageMaker real-time inference endpoint with pre-configured settings optimized on your workload (low value, low latency, throughput, or balanced). No handbook container or occasion configuration required.
Mannequin customization — Tremendous-tune fashions instantly from the console or SDK utilizing methods like low-rank adaptation (LoRA), supervised fine-tuning (SFT), direct desire optimization (DPO), and reinforcement studying with verifiable rewards (RLVR). Serverless customization choices cut back infrastructure administration.

You’ll be able to entry SageMaker JumpStart via the SageMaker Studio console (Fashions → JumpStart fashions), programmatically via the SageMaker Python SDK, or instantly from Hugging Face utilizing the not too long ago launched deep hyperlink integration.

2026 onboarding at a look

From January via September 2026, SageMaker JumpStart onboarded over 100 new fashions throughout 30 distinct suppliers. This represents a big enlargement of the catalog, pushed by the accelerating tempo of open supply mannequin releases and rising buyer demand for alternative throughout mannequin architectures, sizes, and capabilities.

Right here’s a breakdown of the main contributors and a hyperlink to our SageMaker JumpStart web page:

Alibaba (Qwen): 29 fashions spanning reasoning (Qwen 3.8-27B), large-scale mixture-of-experts (Qwen3.5-122B-A10B), agentic workflows (Qwen-AgentWorld-35B-A3B), code technology (Qwen3-Coder-Subsequent, 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: 17 fashions together with the Cosmos3 world basis mannequin household (Tremendous, Edge, Nano), Cosmos Purpose 7B, MSA-Search, Nemotron 3.5 Lightning, Nemotron-3-Tremendous-120B (FP8 and BF16), Nemotron3 Extremely, Nemotron-Nano-12B-v2, Nemotron 3 Nano Omni (multimodal), and Nemotron 3.5 Tremendous VL for vision-language.
Mistral: 7 fashions together with the Ministral-3 household (3B, 8B, 14B in normal and BF16) and Voxtral-Mini-4B-Realtime for speech.
Google: 6 Gemma 4 variants, together with gemma-4-31B-it, E2B-it, E4B-it, 26B-A4B-it, assistant, and NVFP4 quantized.
IBM: 4 fashions together with Granite 4.0 h-micro/small/tiny and granite-speech-4.1-2b for multilingual automated speech recognition (ASR) throughout 6 languages.
Cohere: 4 fashions together with Command A (A100) and Rerank v3.5, v4.0 Quick, and v4.0 Professional for retrieval augmentation.
Others: 33 fashions 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 extra.

Mannequin onboarding highlights

The fashions onboarded in 2026 mirror a number of targeted priorities in how we develop the SageMaker JumpStart catalog. These aren’t solely trade tendencies. They signify the areas the place we see robust buyer demand and vital alternative to assist prospects construct manufacturing AI purposes on SageMaker.

Increasing the NVIDIA open-model portfolio

NVIDIA is one among our vital mannequin suppliers, and in 2026 their contribution prolonged nicely past language fashions. The Cosmos3 world basis fashions, out there in Tremendous, Edge, and Nano tiers, convey video technology, physics simulation, and motion planning right into a single structure for bodily AI purposes equivalent to robotics, autonomous techniques, and digital twins. Mixed with Nemotron fashions spanning from environment friendly edge inference (Nano-12B) to large-scale reasoning (Tremendous-120B), SageMaker JumpStart now provides a broad portfolio of NVIDIA open fashions. If you happen to construct on NVIDIA {hardware}, you’ll be able to deploy these fashions on SageMaker with optimized containers and pre-configured endpoints.

Closing the multimodal hole

Buyer workloads more and more require fashions that course of greater than textual content. In 2026, we prioritized onboarding fashions that deal with imaginative and prescient, speech, and doc understanding natively. NVIDIA Nemotron 3 Nano Omni processes textual content, imaginative and prescient, and speech in a single mannequin. Google Gemma 4 provides vision-language capabilities throughout a number of sizes. IBM granite-speech-4.1-2b handles speech recognition and translation throughout six languages. Mistral Voxtral-Mini helps real-time speech understanding. These additions imply prospects can keep away from chaining separate preprocessing fashions for multimodal workloads.

Making artistic AI accessible

Picture and video technology moved from area of interest to mainstream in 2026. We onboarded FLUX.2 Small Decoder and FLUX.2-klein-base-4B from Black Forest Labs for quick text-to-image technology, Meta Muse-Glimmer-30B for multimodal agentic workflows, and NVIDIA Cosmos3 for video technology. With these fashions, you’ll be able to entry artistic AI capabilities with the identical one-click deployment expertise you employ for language fashions.

Prioritizing environment friendly and quantized fashions

Not each workload requires a frontier-scale mannequin, and many shoppers are targeted on optimizing inference prices as they transfer to manufacturing. We onboarded a number of quantized and compact fashions: Zhipu AI GLM-5.2 FP8 delivers robust efficiency with a diminished reminiscence footprint. JetBrains Mellum2-12B makes use of a mixture-of-experts structure with solely 2.5B energetic parameters for code technology. OpenBMB MiniCPM5-1B packs multimodal capabilities into 1 billion parameters. Microsoft Phi-4-mini-instruct and IBM Granite 4.0 h-tiny show that compact fashions serve manufacturing workloads on occasion varieties like ml.g5.xlarge. These fashions can optimize latency and complete value of possession (TCO) for focused use instances.

Onboarding purpose-built fashions for focused duties

Common-purpose LLMs are versatile, however purpose-built fashions are optimized for particular duties. In 2026, we onboarded Qwen-AgentWorld-35B-A3B for autonomous AI workflows with native tool-calling. GLM-OCR, LightOnOCR, and DeepSeek-OCR deal with doc textual content extraction. Qwen3-Reranker-4B and Cohere Rerank v4.0 enhance retrieval precision in Retrieval Augmented Era (RAG) pipelines. langcache-embed-v3-small gives light-weight embedding for semantic search. Qwen3-ASR and Qwen3-TTS convey speech recognition and synthesis. OpenFold3 helps protein construction prediction for scientific computing. With these fashions, you’ll be able to construct pipelines the place every stage makes use of the specialised software for the job, slightly than forcing a single mannequin to do the whole lot.

Broadening supplier range

Decreasing focus danger in mannequin choice is necessary for enterprise prospects. In 2026, we expanded the catalog to incorporate 30 suppliers, up from a extra concentrated set in prior years. Alibaba Qwen is the only largest contributor with 29 fashions spanning reasoning, coding, speech, and retrieval. Zhipu AI and DeepSeek add aggressive LLMs and specialised OCR. MiniMax and Moonshot AI contribute rising architectures. Cohere brings dependable reranking. BharatGenAI and Kakao develop regional mannequin protection. This breadth provides prospects extra choices for efficiency, value, licensing, and language assist.

SageMaker JumpStart enhancements in 2026

Past mannequin onboarding, SageMaker JumpStart additionally shipped vital enhancements in 2026 that cut back friction throughout the mannequin lifecycle, from discovery to deployment.

Simplified Mannequin Hub on the SageMaker AI console

Beforehand, discovering SageMaker JumpStart fashions required navigating into SageMaker Studio and discovering the SageMaker JumpStart part within the left navigation. The brand new Mannequin Hub expertise surfaces the complete SageMaker JumpStart catalog instantly on the SageMaker AI console touchdown web page. You’ll be able to browse fashions by class and filter by enter/output modality, writer, mannequin dimension, license, and framework. You can too view pattern notebooks and open fashions inside Studio with a single click on. This avoids the multi-step navigation that slowed preliminary exploration.

Pre-configured endpoint deployment choices

Choosing the proper occasion kind, serving container, and optimization settings for a mannequin deployment usually required handbook benchmarking throughout dozens of mixtures. SageMaker JumpStart now provides named deployment configurations, with optimization choices for:

Balanced — A good tradeoff between value, latency, and throughput for general-purpose workloads.
Throughput optimized — Most tokens per second utilizing methods like Speculative Decoding, very best for high-volume batch processing.
Latency optimized — Lowest Time to First Token (TTFT), prioritizing pace for interactive purposes.
Price optimized — Lowest infrastructure value per request, appropriate for workloads that may settle for greater latency.

These configurations embody Neuron-optimized setups for AWS Inferentia and AWS Trainium cases, decreasing deployment complexity and infrastructure prices. You’ll be able to deploy a mannequin with a single click on and the best optimization profile, slightly than spending weeks benchmarking configurations manually.

Hugging Face deep hyperlink integration

A brand new integration with Hugging Face connects mannequin discovery on to SageMaker Studio, for onboarded fashions. When searching fashions on Hugging Face, you’ll be able to select Deploy on SageMaker AI or Customise and Deploy on SageMaker AI. You land instantly in a configured Studio setting with the mannequin pre-loaded, permissions pre-configured, and GPU quota visibility inbuilt. This avoids the necessity to arrange AWS Id and Entry Administration (IAM) roles, create a website, or configure permissions earlier than experimenting.

Getting began

There are 3 ways to get began with SageMaker JumpStart:

SageMaker Studio console — Open SageMaker Studio, navigate to Fashions → JumpStart fashions, and browse the catalog. Select a mannequin to view its particulars, then select Deploy or Customise.
SageMaker Python SDK — Deploy fashions programmatically utilizing the SageMaker Python SDK. Go to the SageMaker JumpStart mannequin catalog for SDK examples and deployment guides.

Hugging Face deep hyperlink — Browse the HuggingFace Mannequin Hub and select Deploy on SageMaker AI to land instantly in Studio for SageMaker JumpStart onboarded fashions.

Every mannequin on SageMaker JumpStart consists of deployment configurations, pattern notebooks, and documentation. For a hands-on walkthrough, see our associated weblog publish on “Deploy and consider fashions with Amazon SageMaker JumpStart”.

Conclusion

The primary three quarters of 2026 have been a interval of great progress for Amazon SageMaker JumpStart. With over 100 new fashions onboarded from 30 suppliers, the catalog now spans a variety of capabilities. These embody frontier reasoning fashions, multimodal vision-language fashions, artistic AI for picture and video technology, speech recognition and synthesis, optical character recognition (OCR) and doc understanding, embedding and retrieval fashions for RAG, agentic workflow fashions, and specialised scientific computing fashions.

These onboarding priorities mirror the place we see the strongest buyer demand and the best alternative to speed up AI adoption on AWS. They embody increasing the portfolio, closing the multimodal hole, making artistic AI accessible, prioritizing effectivity and quantization, onboarding purpose-built fashions, and broadening supplier range.

Mixed with the enhancements like simplified Mannequin Hub, pre-configured deployment choices, and the Hugging Face deep hyperlink integration, SageMaker JumpStart continues to decrease the barrier between mannequin discovery and manufacturing deployment. Go to the SageMaker JumpStart mannequin catalog to discover the complete lineup.