模型後訓練／開發平台
Prime Intellect Converges RLVR, LoRA, and Deployment for Nemotron 3 Nano into a Managed Training Workflow
NVIDIA has released a reproducible managed post-training workflow that lets developers use the Prime CLI to run 100 steps of RLVR on a 30B MoE model, then download or deploy the resulting LoRA. In the demonstration, performance rose from 21.9% to 90.6% across 32 math problems, though the sample was small and both the tasks and rewards were tightly controlled.
NVIDIA and Prime Intellect released an end-to-end customization example for Nemotron 3 Nano on July 23, bringing baseline evaluation, reinforcement learning with verifiable rewards, LoRA artifacts, and inference deployment into a single managed workflow. Developers use the Prime CLI to create a workspace, pin environment versions, and select a hosted model without having to configure a multi-GPU training cluster themselves.
The example uses NVIDIA-Nemotron-3-Nano-30B-A3B-BF16. It has 30 billion total parameters, with roughly 3.5 billion parameters activated per token, and combines 23 Mamba-2 layers, 23 MoE layers, and six attention layers. Each MoE layer contains 128 routed experts and one shared expert, with six experts activated. The training environment requires the model to solve problems using Python, NumPy, SymPy, or SciPy. A rule-based grader assigns binary rewards to the outputs, eliminating the need for another large model to serve as a judge.
The official configuration runs for 100 steps, samples each problem eight times, uses a batch of 32 rollouts, sets the learning rate to 2e-5, and limits each solution attempt to five assistant turns. After training, the platform produces a downloadable LoRA adapter, which can also be attached to a hosted inference endpoint for the original model. NVIDIA reports that the average reward on the same 32 held-out problems increased from 0.219 to 0.906—equivalent to an increase from seven correct answers to 29—at a training cost of less than five dollars.
The technical significance of this release lies not in a new model, but in the commercialization of the RLVR experimentation loop: the environment, model version, rollouts, reward curves, and adapter can all be tracked. However, 32 problems are insufficient to demonstrate an improvement in general mathematical ability, and the formatting reward may primarily teach the model to stop misusing tools and to return answers in `boxed` format. Before adoption, teams should add a larger, unseen test set, check for regressions in existing capabilities, and compare the cost-effectiveness of LoRA, supervised fine-tuning, and prompt-only corrections.