Customize Nemotron 3 Nano with Hosted RLVR in Five Minutes

NVIDIA and Prime Intellect Lab published a tutorial on July 23, 2026, showing how to customize Nemotron 3 Nano using hosted reinforcement learning with verifiable rewards (RLVR). The workflow takes about five minutes to set up locally, with the actual training and evaluation running on Prime Intellect Lab infrastructure, so developers do not need to manage a local GPU cluster.

The tutorial follows a baseline, train, and reevaluate flow. On 32 held-out math problems from the Python Math environment, Nemotron 3 Nano started with 21.9% accuracy and reached 90.6% after 100 training steps, with total spend under $5. The same configuration applies to Nemotron 3 Super and Nemotron 3 Ultra by changing the model identifier in the training config.

The three-step workflow

Step one establishes a baseline. After installing the Prime CLI and authenticating, the developer creates a workspace and pins the math-python environment package version 0.1.10. A baseline evaluation runs the model on 32 examples with a binary reward: 1 for a correct boxed answer, 0 for incorrect or missing. The tutorial shows an example where the model fails because it exhausts five assistant turns calling unsupported tools without producing a final response.

Step two launches training. The config file specifies the model, loss type, 100 maximum steps, batch size 32, rollouts per example 8, learning rate 2e-5, and LoRA alpha 16. The trainer samples eight attempts per task to get a comparative signal within each rollout group. The CLI prints the parsed configuration, current price, and environment action status before asking for confirmation, then returns a run ID and dashboard URL. The dashboard and CLI expose reward curves, reward distributions, and individual rollouts for inspection.

Step three checks the result. When the run completes, Prime Intellect Lab uploads the final LoRA adapter. The developer confirms the adapter exists and is ready to deploy, then requests deployment and watches its state move through DEPLOYING to DEPLOYED. The tutorial recommends comparing baseline and adapted rollouts side by side, checking whether the model produces correct boxed answers and uses tools appropriately.

What customization means here

Customization changes a model's trainable parameters so it becomes more likely to succeed at target tasks. The tutorial focuses on RLVR in the Python Math environment, which uses Python tools including numpy, sympy, and scipy for calculations. The environment defaults to 100 turns, but this experiment caps each rollout at five assistant turns, penalizing repeated tool calls without a final response.

Open weights are only part of the equation, according to the tutorial. Data, evaluation code, and training recipes also determine what developers can understand, reproduce, and deploy. The Nemotron 3 family is published with open model resources including weights, data, and recipes.

Prerequisites include a development environment with curl and a supported Python 3 installation, a Prime Intellect account with Hosted Training and billing configured, and internet access for the Prime CLI, environment package, hosted model, and adapter deployment. The tutorial notes that model availability and pricing change over time, so the live CLI catalog should be treated as the source of truth rather than copying an identifier or price from a static article.