AWS and NVIDIA launch serverless fine-tuning for Nemotron 3 models

AWS and NVIDIA have integrated serverless fine-tuning for Nemotron 3 models into Amazon SageMaker AI, letting enterprises adapt open-weight LLMs without provisioning infrastructure. The Nemotron 3 family uses a hybrid Mamba-Transformer MoE architecture that interleaves Mamba-2 layers, Transformer attention, and LatentMoE—activating only a fraction of parameters per forward pass (e.g., 12B of 120B for Super) for high throughput on up to 1M-token contexts. Supported techniques include supervised fine-tuning (SFT), reinforcement learning with verifiable rewards (RLVR), and reinforcement learning from AI feedback (RLAIF).

The two models available are Nemotron 3 Nano (30B total, 3B active) for cost-efficient high-volume multi-agent workloads, and Nemotron 3 Super (120B total, 12B active) for complex reasoning, coding, and long-context analysis. Both were trained via multi-environment RL in NeMo Gym, achieving 4x throughput improvement over prior generations. By fine-tuning on proprietary data, organizations encode unique intelligence into the model weights, turning general-purpose AI into a competitive advantage—often matching larger closed models at a fraction of the compute and cost, all within a secure, private environment.