This page is for version 0.11.
For other versions, use one of these documentation indexes:
- Latest (default): https://docs.nvidia.com/nemo-helix/latest/llms.txt
- 0.11: https://docs.nvidia.com/nemo-helix/v0.11/llms.txt
- 0.10: https://docs.nvidia.com/nemo-helix/v0.10/llms.txt
- 0.9: https://docs.nvidia.com/nemo-helix/v0.9/llms.txt
- 0.8: https://docs.nvidia.com/nemo-helix/v0.8/llms.txt
- 0.7: https://docs.nvidia.com/nemo-helix/v0.7/llms.txt
- 0.6: https://docs.nvidia.com/nemo-helix/v0.6/llms.txt
- 0.5: https://docs.nvidia.com/nemo-helix/v0.5/llms.txt
This matrix defines the supported OSS targets for NeMo Helix. It covers the
Python package, CLI, SDK, local services started by nemo setup, and
self-managed Kubernetes deployments installed with NeMo Helix Helm chart.
NeMo Libraries and Integrations
Versions are generated from this release’s dependency lockfiles and image build pins. Optional libraries are not installed in every configuration.
| Library / distribution | Version / source revision | Source of truth |
|---|---|---|
data-designer | 0.9.7 | uv.lock |
data-designer-config | 0.9.7 | uv.lock |
data-designer-engine | 0.9.7 | uv.lock |
data-designer-retrieval-sdg | 0.2.1 | uv.lock |
garak | 0.17.1 | docker/Dockerfile.nhx-garak-tasks |
nemo-anonymizer | 0.4.0 | uv.lock |
nemo-automodel | r0.7.0 (Git 48c88e16798246e8c7f1b9ea3974bfc488bd0849) | docker/automodel/Dockerfile.nhx-automodel-base |
nemo-compass | 0.2.0 | uv.lock |
nemo-fabric | 0.6.0 | uv.lock |
nemo-fabric-adapter-catalog | 0.6.0 | uv.lock |
nemo-fabric-adapter-contract | 0.6.0 | uv.lock |
nemo-fabric-adapters-claude | 0.6.0 | uv.lock |
nemo-fabric-adapters-codex | 0.6.0 | uv.lock |
nemo-fabric-adapters-common | 0.6.0 | uv.lock |
nemo-fabric-adapters-deepagents | 0.6.0 | uv.lock |
nemo-fabric-adapters-hermes | 0.6.0 | uv.lock |
nemo-fabric-adapters-mini-swe-agent | 0.6.0 | uv.lock |
nemo-fabric-adapters-nooa | 0.6.0 | uv.lock |
nemo-fabric-adapters-remote-agent | 0.6.0 | uv.lock |
nemo-fabric-collector | 0.6.0 | uv.lock |
nemo-fabric-evidence | 0.6.0 | uv.lock |
nemo-fabric-runtime | 0.6.0 | uv.lock |
nemo-gym | Submodule of NeMo-RL at f34b91df909e4d9f8fa2faec66ce95fb746a07fa | docker-bake.hcl |
nemo-gym | 0.7.0 | docker/locks/nhx-gym-host/uv.lock |
nemo-gym | 0.7.0 | docker/locks/nhx-gym-tasks/uv.lock |
nemo-helix-tavily-plugin | 0.2.0 | uv.lock |
nemo-relay | 0.11.1 | docker/locks/nhx-agent-hardener-tasks/uv.lock |
nemo-relay | 0.11.0 | uv.lock |
nemo-relay-cli-bin | 0.11.0 | uv.lock |
nemo-rl | Git f34b91df909e4d9f8fa2faec66ce95fb746a07fa | docker-bake.hcl |
nemo-safe-synthesizer | 0.1.17 | uv.lock |
nemo-switchyard | 0.3.0 | uv.lock |
nemoguardrails | 0.24.1 | uv.lock |
nvidia-agent-hardener | 0.0.18 | docker/locks/nhx-agent-hardener-tasks/uv.lock |
nvidia-nat-atif | 1.9.0a20260811 | uv.lock |
nvidia-nat-config-optimizer | 1.9.0a20260811 | uv.lock |
nvidia-nat-core | 1.9.0a20260811 | uv.lock |
nvidia-nat-eval | 1.9.0a20260811 | uv.lock |
nvidia-nat-langchain | 1.9.0a20260811 | uv.lock |
nvidia-nat-opentelemetry | 1.9.0a20260811 | uv.lock |
Host Platforms
| Area | Supported | Notes |
|---|---|---|
| Deployment mode | nemo setup local install; Helm chart on Kubernetes | nemo setup starts local services, registers a model provider, and configures the CLI and SDK. The Helm chart is the supported path for self-managed Kubernetes clusters. |
| Python | 3.12, 3.13, 3.14 | Use an isolated virtual environment. Python 3.12 is the lowest supported Python version and Python 3.14 is the highest supported Python version. |
| Linux | Ubuntu 22.04 LTS, Ubuntu 24.04 LTS, RHEL 9, Rocky Linux 9, Debian 12 | Supported for CLI, SDK, local services, and NVIDIA GPU workloads. |
| macOS | macOS Tahoe 26 and macOS Sequoia 15 | Supported for CLI, SDK, local services, and cloud/provider workflows. Local NVIDIA GPU workloads are not supported on macOS. |
| Architecture | x86_64 Linux; Apple Silicon and Intel macOS | NVIDIA GPU workloads require x86_64 Linux. |
| Windows and WSL | Not supported for the OSS local install path | Use a supported Linux or macOS host for the local install path. |
| Kubernetes clusters | minikube, kind, EKS, AKS, GKE, OKE, OpenShift, and on-prem Kubernetes | See Install NeMo Helix with Helm for prerequisites and cluster-specific notes. |
Local Runtime
| Area | Supported | Notes |
|---|---|---|
| Package installation | uv>=0.10.10 recommended for source checkouts; pip supported for published packages | The one-line installer sets up uv automatically. Published packages can also be installed with pip install "nemo-helix[all]". |
| Shell | Bash or zsh | Used by the installer, setup commands, and documented examples. |
| Browser | Current Chrome, Edge, Firefox, or Safari | Required for the Studio UI and generated documentation. |
GPU Support
| Area | Supported | Notes |
|---|---|---|
| CUDA | CUDA 13 or later | GPU workloads require CUDA 13 or later. |
| NVIDIA driver | CUDA 13-capable driver, R580 branch or later | See the CUDA 13.0 release notes for exact driver minimums by operating system. |
| GPU memory | 40 GB VRAM minimum; 80 GB VRAM recommended | Smaller GPUs can run provider or client workflows, but local model and GPU-accelerated jobs require larger GPUs. |
| Recommended GPUs | NVIDIA A100 80GB, H100 80GB, B200 180GB | Use comparable data center GPUs that meet the CUDA and memory requirements for the model or job. |
| GPU platform | Linux x86_64 | macOS and Windows are not supported for local NVIDIA GPU workloads. |
Resource Profiles
| Profile | CPU and memory | Storage | GPU |
|---|---|---|---|
| CLI, SDK, and hosted-provider setup | 8 GB RAM minimum | 16 GB free disk space | Not required |
| Local GPU workloads | 16 GB RAM minimum; 32 GB recommended | Additional tens to hundreds of GB for model artifacts and job outputs | 40 GB VRAM minimum; 80 GB recommended |
Model Providers
| Provider | Support | Notes |
|---|---|---|
| NVIDIA Build | Supported | Configure with NVIDIA_API_KEY. |
| OpenAI | Supported | Configure with OPENAI_API_KEY. |
| Anthropic | Supported | Configure with ANTHROPIC_API_KEY. |
| Google Gemini | Supported | Configure with GEMINI_API_KEY. |
| Ollama | Supported for local provider workflows | Ollama must be installed and running before setup. |
| Custom OpenAI-compatible endpoint | Supported | Configure NEMO_DEFAULT_INFERENCE_BASE_URL and NEMO_DEFAULT_INFERENCE_KEY. |
Out of Scope
The following are not part of the OSS support matrix:
- Native Windows local install.
- Docker Compose deployment guides.
- WSL-based local install or GPU validation.
- macOS local GPU workloads.
- Python earlier than 3.12 or Python 3.15 and later.
- CUDA versions earlier than 13 for GPU workloads.