NeMo Studio
Model deployment and fine-tuning in NVIDIA NeMo Studio: what it offers
NVIDIA NeMo Studio, the web interface of NeMo Helix, can deploy a model and fine-tune one. Deployment starts a vLLM or NVIDIA NIM container from a Hugging Face repository, a workspace model or an NGC image, as a single replica, and it does not use NVIDIA Dynamo. Fine-tuning is broader: three training backends, LoRA or full weights, distillation and reinforcement learning. This is from the screens of release 0.9.1 and the open-source code on 14 and 15 November 2026. We did not run a deployment or a training job, because our test machine had no GPU.
Model deployment
| Choice | Options |
|---|---|
| Where the model comes from | A Hugging Face repository, a model or weights already in the workspace, or an NGC NIM container image |
| Engine | vLLM from a default image, or NIM. The code has a third engine that runs any container |
| GPUs | A single count per deployment |
| LoRA | A switch that serves adapters trained against the base model alongside it |
| Advanced | Disk size for the weights and environment variables |
- A deployment becomes a Docker container or a Kubernetes Deployment.
- The Kubernetes path sets one replica. We found no autoscaling, scale to zero or multi-node serving in the model service.
- Deployed models are reached through an inference gateway and named virtual models.
- NVIDIA Dynamo does not appear in the NeMo Helix source. Serving at scale is left to other NVIDIA software.
Fine-tuning
There are three ways in: describe the goal and let an assistant draft the settings, start from a template, or fill in the full form. The templates in release 0.9.1 are text-to-SQL recipes on Nemotron models.
| Choice | Options |
|---|---|
| Training backend | Automodel for multi-GPU distributed training, Unsloth for a single GPU with 4-bit quantisation, RL for reinforcement learning |
| Fine-tuning type | LoRA, LoRA merged into the base weights, or full weights |
| Training type | Supervised fine-tuning, or distillation from a larger teacher model |
| Reinforcement learning | DPO on preference pairs, or GRPO with a reward environment |
| Data | A JSONL dataset split into training and validation |
| Parameters | Epochs, learning rate, batch size, sequence length, LoRA rank, alpha and dropout |
| Compute | Nodes and GPUs per node |
- Training runs as a container job on a GPU cluster.
- A LoRA adapter trained against a base model is picked up by a deployment of that model.
How to read this
- Fine-tuning is the fuller of the two features. It suits a team that already has NVIDIA GPUs and wants several training methods behind one form.
- Deployment is a convenient way to serve one model for testing. It is not a production serving layer.
- You enter the hardware yourself. The form does not tell you whether a model fits a GPU or what a run will cost.
Where Katonic fits
Katonic covers the same two jobs with a different emphasis. Its fine-tuning wizard offers LoRA and QLoRA, checks whether the model fits the GPU, estimates the cost before launch, and carries the result through comparison to deployment. Its model deployment is built for many models and many teams. NeMo Studio offers more training methods today, including distillation and reinforcement learning.
Read next
- What is NVIDIA NeMo Helix? →
- NVIDIA NeMo Helix, Relay and Fabric: field guide →
- How to install NVIDIA NeMo Helix →
- NeMo Helix on GitHub →
- NeMo Studio documentation →
Questions and answers
Can NVIDIA NeMo Studio deploy a model?
Yes. It starts a vLLM or NIM container from a Hugging Face repository, a workspace model or an NGC image, as a single replica.
Does NeMo Helix use NVIDIA Dynamo?
No. Dynamo does not appear in the NeMo Helix source as of November 2026. Model deployment uses vLLM and NIM containers.
Can NVIDIA NeMo Studio fine-tune a model?
Yes. It offers three training backends, LoRA or full-weight tuning, distillation and reinforcement learning.
Which fine-tuning backends does NeMo Studio have?
Automodel for multi-GPU distributed training, Unsloth for a single GPU, and RL for reinforcement learning with DPO or GRPO.
Does NeMo Studio autoscale model deployments?
We found no autoscaling in the model service. The Kubernetes path sets one replica.
Did you run these features?
No. Our test machine had no GPU. This page is from the screens and the open-source code.