# AWS SageMaker AI Adds Serverless Fine-Tuning for NVIDIA Nemotron 3.5 Lightning > Source: > Published: 2026-09-17 01:12:05+00:00 **September 17, 2026, (Inside AI) —** Amazon Web Services has added serverless model customization for NVIDIA's Nemotron 3.5 Lightning to SageMaker AI, allowing enterprises to fine-tune the open-weight model without managing GPU clusters. The service supports supervised fine-tuning (SFT), Direct Preference Optimization (DPO), and reinforcement fine-tuning (RFT), according to an AWS announcement. The move targets a common enterprise pain point: adapting large foundation models to proprietary data while controlling cost and latency. Nemotron 3.5 Lightning uses a hybrid Mixture-of-Experts architecture with **3 billion** active parameters and **30 billion** total parameters, making it efficient for inference but still capable of domain-specific tasks after customization. Serverless customization means SageMaker AI provisions infrastructure and orchestrates training automatically. Customers pay only for what they use, avoiding idle cluster costs. The service is available in **US East (N. Virginia)**, **US West (Oregon)**, **Asia Pacific (Tokyo)**, and **Europe (Ireland)**. Enterprises can now tailor Nemotron 3.5 Lightning with labeled data via SFT for accuracy, preference data via DPO for tone alignment, or reward signals via RFT for new task performance. This approach aims to let smaller, right-sized models match frontier-model quality on specific tasks, reducing both cost and latency. **Read:** **Abu Dhabi's open-source AI models** ## Why Serverless Fine-Tuning Matters Now The launch reflects a broader shift in enterprise AI: moving from monolithic foundation models to smaller, customized models that run efficiently. AWS competes with similar offerings from Google Cloud and Microsoft Azure, but the serverless angle removes a key barrier for teams without deep MLOps expertise. NVIDIA's Nemotron 3.5 Lightning is one of the latest open-weight models from the chipmaker, designed for high throughput and low latency. Its Mixture-of-Experts design activates only a fraction of parameters per token, which cuts inference costs. Fine-tuning such a model traditionally required significant GPU resources, but serverless customization abstracts that away. "With serverless customization, SageMaker AI handles all infrastructure provisioning and training orchestration, so you can focus on your data and evaluation rather than cluster management, and only pay for what you use," an AWS spokesperson said. The service supports programmatic access via the SageMaker Python SDK and a visual interface in Amazon SageMaker Studio. Users navigate to the Models page to launch a customization job. Documentation is available for detailed steps. Industry analysts note that fine-tuning open-weight models has become a competitive differentiator. Companies like Hugging Face and Databricks offer similar capabilities, but AWS's integration with its cloud ecosystem may appeal to existing customers. The serverless model also aligns with the trend of "right-sizing" AI: using smaller models for specific tasks rather than relying on massive general-purpose models. However, challenges remain. Fine-tuning still requires high-quality labeled data, which many enterprises lack. DPO and RFT demand preference or reward signals that can be costly to collect. AWS's service simplifies infrastructure, but data preparation remains a bottleneck. Pricing details were not disclosed, but AWS said customers pay only for what they use. This contrasts with provisioned cluster pricing, where idle time incurs costs. The serverless approach could lower the entry barrier for smaller teams. **Read:** **Qwen3.8-Flash lower training costs** NVIDIA's Nemotron 3.5 Lightning is available for deployment on SageMaker AI as well as customization. The model's open weights allow enterprises to inspect and modify it, addressing data sovereignty concerns. Looking ahead, AWS is likely to expand serverless customization to more models. The company has been adding support for various open-weight models, including Meta's Llama and Mistral AI's offerings. This move strengthens its position in the generative AI platform market. For now, enterprises in the four available regions can start customizing Nemotron 3.5 Lightning. The service is live, and AWS encourages users to explore the documentation to get started.