How to fine-tune Nemotron 3 Nano (LoRA, QLoRA)

This guide covers adapting the Nemotron 3 family with LoRA or QLoRA. Estimate cost with the training cost calculator and see fine-tuning for the overview.

Which checkpoint

• The Base checkpoint (NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16) is trained with next-token prediction. Its card calls it "a good starting point for instruction fine-tuning" for developers building instruction-following models.
• The main BF16 checkpoint is the post-trained one with reasoning and tool calling. Start from it to keep that behaviour.

Training framework

The post-trained card recommends the NeMo Framework for fine-tuning. The cards give no LoRA target-module list, so inspect the module names yourself. The model is a Mamba2-Transformer hybrid MoE (23 MoE, 23 Mamba-2 and 6 attention layers), so confirm your trainer supports those layers and the custom code (trust_remote_code).

Chat template

Use the model's template. Reasoning is controlled by enable_thinking in apply_chat_template() and is on by default. If you train a non-reasoning variant, build your data with enable_thinking=False so the format matches inference.

License

Both checkpoints use the NVIDIA Nemotron Open Model License. Per the license text:

• You may create and distribute derivative works, and commercial use is permitted.
• Redistribution requires a copy of the license, retained notices, a NOTICE file, and the statement "Licensed by NVIDIA Corporation under the NVIDIA Nemotron Model License."
• Patent or copyright litigation against NVIDIA over the work terminates your license.
• NVIDIA does not claim ownership of outputs.