AWS Adds NVIDIA Nemotron 3 Nano (30B) and Super (120B) to SageMaker Serverless Fine-Tuning

Amazon SageMaker AI now supports serverless fine-tuning for NVIDIA Nemotron 3 models, starting with Nemotron 3 Nano (30B total parameters, 3B active) and Nemotron 3 Super (120B total parameters, 12B active). The integration eliminates infrastructure provisioning and management for enterprises fine-tuning these models on domain-specific data.

Architecture and Capabilities

Nemotron 3 uses a hybrid Mamba-Transformer Mixture-of-Experts (MoE) architecture with native support for up to 1M-token context lengths. The architecture interleaves three layer types: Mamba-2 layers for linear-time sequence processing, Transformer attention layers for associative recall, and Latent Mixture-of-Experts (LatentMoE) layers that compress tokens before routing to specialized experts.

The MoE design activates only a fraction of total parameters per forward pass—12B of 120B in the Super variant. According to NVIDIA, Nemotron 3 Nano achieves 4x higher throughput than its predecessor Nemotron 2 Nano while maintaining strong performance on coding and reasoning tasks.

Fine-Tuning Techniques

SageMaker AI's serverless customization supports three techniques:

Supervised Fine-Tuning (SFT): Uses labeled input-output pairs to teach new behaviors. Best for domain Q&A pairs, formatted tool calls, and task-specific instruction completions.

Reinforcement Learning with Verifiable Rewards (RLVR): Optimizes model behavior against a reward signal. The model generates multiple candidate responses per prompt, a reward function scores them, and the model updates its policy accordingly. Suited for tasks with verifiable objectives like code correctness or format compliance.

Reinforcement Learning from AI Feedback (RLAIF): Uses a separate AI model to evaluate outputs and provide feedback signals for policy improvement. Used for aligning model tone, helpfulness, and safety without human-labeled reward data.

Model Variants

Nemotron 3 Nano 30B: Optimized for high-volume, multi-agent workloads where cost and latency matter. The 3B active parameter footprint targets specialized tasks requiring high compute efficiency.

Nemotron 3 Super 120B: Designed for complex multi-agent AI and reasoning tasks requiring more capacity. According to NVIDIA, it performs well at reasoning, coding, and long-context analysis while maintaining efficiency for continuous operation at scale.

Both models were trained using multi-environment reinforcement learning through NeMo Gym, which NVIDIA claims aligns them to multi-step agentic tasks across coding, reasoning, and long-context analysis.

Implementation

Users can access serverless customization through the SageMaker Studio console or programmatically via the SageMaker Python SDK. The service handles infrastructure provisioning, training orchestration, checkpointing, and fault tolerance. Pricing is usage-based with no upfront infrastructure costs.

AWS positions the integration as part of SageMaker AI's broader catalog of open-weight models available for serverless customization. The company states that fine-tuning smaller open-weight models on targeted tasks often matches or exceeds the performance of larger proprietary models while keeping sensitive data within private infrastructure.

What This Means

This integration gives AWS customers access to NVIDIA's latest hybrid architecture models without infrastructure overhead. The 1M token context window and MoE efficiency make these models viable for enterprise use cases like IT automation and multi-agent systems. The availability of three fine-tuning techniques—SFT, RLVR, and RLAIF—allows teams to optimize for different objectives: teaching specific behaviors, maximizing verifiable metrics, or aligning subjective qualities. For organizations already on AWS, serverless customization removes a barrier to deploying specialized models in production.