> This page is for version 0.7.0. > For other versions, use one of these documentation indexes: > - Latest (default): https://docs.nvidia.com/nemo-helix/latest/llms.txt > - 0.8.0: https://docs.nvidia.com/nemo-helix/v0.8.0/llms.txt > - 0.7.1: https://docs.nvidia.com/nemo-helix/v0.7.1/llms.txt > - 0.7.0: https://docs.nvidia.com/nemo-helix/v0.7.0/llms.txt > - 0.6.0: https://docs.nvidia.com/nemo-helix/v0.6.0/llms.txt > - 0.5.1: https://docs.nvidia.com/nemo-helix/v0.5.1/llms.txt > - 0.5.0: https://docs.nvidia.com/nemo-helix/v0.5.0/llms.txt > - 0.4.0: https://docs.nvidia.com/nemo-helix/v0.4.0/llms.txt > - 0.3.0: https://docs.nvidia.com/nemo-helix/v0.3.0/llms.txt > - 0.2.0: https://docs.nvidia.com/nemo-helix/v0.2.0/llms.txt > For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.nvidia.com/nemo-helix/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/nemo-helix/_mcp/server. # v0.7.0 - 2026-10-06 NeMo Helix v0.7.0 completes the product rebrand to NeMo Helix across the repository, packages, containers, and Helm chart, and builds on the 0.6 agent, evaluator, and Studio workflows with strategy-based agent optimization, one-step agent redeploys, hardened access-key and identity provider support, and deeper Studio agent-onboarding and Customizer workflows. ## Highlights * **NeMo Helix rebrand.** The repository, Python packages, container images, and Helm chart complete their rename to NeMo Helix, including the remaining legacy container image and Helm chart names. * **Strategy-based agent optimization.** `nemo agents optimize` now exposes `list-strategies`, `run-strategy`, and `prepare-fileset` subcommands so an installed optimization strategy can be resolved, validated, and submitted as a Fabric-backed job, alongside a new `nemo agents redeploy` command for updating a deployed agent from a new config in one step. * **Hardened access and identity.** Scoped access keys can be rotated with a dual-active grace period, created with explicit scopes, and resolved to a stable account identity. Auth adds a ZITADEL Kubernetes identity provider reference and configurable CLI device-flow and Studio bearer-token behavior. * **Deeper Studio agent and Customizer workflows.** Studio can create Fabric agents from a GitHub repository, from an agent seen in imported traces, or from uploaded loose files, package an agent image directly from Studio, and launch agent HPO studies from the Optimize button. Deployments and the Inference Provider feature flags are enabled by default, and Customizer gets a landing page and combined fileset/model/dataset creation. * **Broader Fabric evaluation coverage.** Evaluator can run Fabric harnesses, including deepagents on Nemotron, against Harbor tasks that ship no Python, submit Gym runners with secret references and placement, and submit agent evaluations from the async plugin SDK. ## What's included ### Agents and Fabric * `nemo agents optimize` adds `list-strategies` (list installed optimization strategies), `run-strategy --strategy ` (submit a Fabric-backed optimization job for the named strategy, for example `legacy`), and `prepare-fileset` (validate an optimize bundle and stage it as a fileset, with an optional `--check-models` model resolution check). * `nemo agents redeploy --agent-config ` redeploys an agent from a new config in one command instead of separate update and restart steps. * Agent deployments record which GitHub commit they are running. * Execute-agent job configuration accepts an explicit image field alongside other agent config fields. ### Agent Evaluation * Evaluator can run a NeMo Fabric deepagents harness on Nemotron inside Harbor, and can run Fabric harnesses on Harbor tasks that ship no Python. * Gym runner submissions can include secret references and a placement target, and agent evaluations can be submitted from the async plugin SDK. * Fabric sandbox and Relay evaluation trials now capture their own OTLP trace, so runs using those Harbor targets show up alongside other traced trials. * Harbor agent evaluation credentials are supplied through env-secrets instead of inline configuration, keeping credentials out of evaluation config files. * Intake can import Gym rollout trajectories. * The sandboxed Gym runtime now vendors its Hugging Face cache into its writable work directory, so sandboxed Gym tasks no longer need external Hugging Face access to run. ### Studio * Studio can create a Fabric agent from a GitHub repository (recording its commit), from an agent seen in imported traces, or by uploading loose files instead of only a whole directory, and can package an agent image directly from Studio. * The Deployments and Inference Provider feature flags are enabled by default, the Create Deployment wizard moved to its own page, and Studio offers deploy CTAs for custom models, adapters, and runs that emit their own model when a model isn't already being served. * Agent details pages add an Insights analysis config panel driven by trace-intel, an Import Traces modal, and contextual CTAs. * The Optimize button can launch agent HPO studies, and the Traces list can be filtered by agent. * Customizer gets a landing page, combined fileset upload with model and dataset creation, deploy-the-base-model support for adapter fine-tuning jobs, and clearer chart series on customization detail pages. * The Run Evaluation form gains explicit steps and a control for how many rows run in parallel. * Navigation, header state, and chevron behavior are more consistent across Studio surfaces. ### Customizer and Data Designer * Automodel training jobs can auto-deploy after completion, and RL training jobs can auto-deploy the resulting model once a job finishes. * Data Designer adds a `check-models` command that validates a configuration's models against the platform before use. ### Platform, CLI, and Deployment * Auth adds `nemo auth access-keys rotate`, which mints a successor key and keeps the prior key active for a configurable grace period, scopes on access-key creation, stable account identity resolution, a ZITADEL Kubernetes identity provider reference, and configurable CLI device-flow and Studio bearer-token behavior. * `nemo chat` uses the default model chosen during `nemo setup` (or `NEMO_DEFAULT_MODEL`) when no `--model` is given, and `nemo setup` creates a sample workspace and agent as part of the setup flow. * The CLI inference providers list shows provider status, and completed CLI customization submissions link to their Studio job details page. * Models reject a provider URL with an unsupported scheme at create time, and the inference gateway routes LoRA adapters through their base model's virtual model instead of a separate route. * Jobs across agents, evaluator, anonymizer, data designer, and safe synthesizer now report token usage as part of their job record. ## Maintenance * The CLI's typed-client plumbing no longer depends on Stainless-generated code. * Fabric-backed agent workflows upgrade to Fabric 0.3.0. ## Install For fresh local installs, see [Setup](/documentation/get-started/setup). It covers the recommended PyPI install, source-checkout prerequisites, Flox activation, the system-toolchain fallback, and provider configuration. For self-managed Kubernetes, start with [Install NeMo Helix Helm Chart](/documentation/kubernetes-deployment/setup/helm/install). If you are updating pinned chart or image references from an earlier release, see the upgrade steps below. ## Upgrade from v0.6.x From an existing local checkout: ```bash git fetch git checkout main make bootstrap flox -q activate ``` Without Flox, install the system toolchain versions printed by `make toolchain-versions`, plus a C compiler, then run `make TOOLCHAIN=system bootstrap` followed by `source .venv/bin/activate`. Run setup after updating the checkout: ```bash nemo setup ``` After setup, restart local services before using CLI, SDK, Studio, or plugin workflows against the upgraded checkout. If you deploy through Helm or reference container images directly, update chart and image references from the legacy product names to the NeMo Helix chart and image names before redeploying. If you use `nemo agents optimize`, move existing automation from the prior single-path optimize flow to `nemo agents optimize run-strategy --strategy legacy` (or another installed strategy), staging remote bundles first with `nemo agents optimize prepare-fileset`. If you have platform-owned agent intent packages, use the `nemo-ethos` coding-agent skill to write `ETHOS.md`. After the user approves `ETHOS.md`, remove the legacy spec package before using newer optimization workflows. ## Compatibility * Python 3.12, 3.13, or 3.14 * macOS and Linux for local CLI, SDK, Studio, and hosted-provider workflows * Linux x86\_64 for local NVIDIA GPU workloads * CUDA 13-capable NVIDIA drivers for local GPU workloads * Self-managed Kubernetes clusters deployed with Helm * Docker for local services, Docker-backed platform jobs, local ClickHouse, Docker Sandboxes, local model-serving workflows, and Gym task images * Kubernetes/Ray runtime for NeMo-RL DPO and GRPO customization jobs * OpenSandbox and a job-storage PVC for sandboxed Gym and GRPO reward environments * ClickHouse for Intake trace storage and Experiments rollups * NVIDIA GPU access for local training, model serving, RL training, and GPU-backed synthetic data workflows * Node 22.18.0+ for Studio assets * Platform API and `nemo-helix` Python SDK `0.7.0` ## Current constraints * **Self-managed scope.** v0.7.0 documents local setup and user-managed Kubernetes deployment. It is not a managed hosted-service release. * **Research-preview optimizer.** Analyst and Experimentalist are early agentic workflows. They are useful for guided optimization, but quality and autonomy are still research-preview and require developer review. * **Sandboxed Gym runtime.** Sandboxed Gym evaluation requires Kubernetes or a configured sandbox-capable environment, OpenSandbox setup for cluster runs, and task images that include the environment dependencies. * **Retrieval evaluation.** Retrieval evaluation requires the relevant BEIR data or retrieval corpus inputs and configured embedding or ranking model access. * **GRPO runtime.** GRPO jobs require Kubernetes/Ray execution and sandboxed Gym configuration. They do not have a local Docker fallback. * **GRPO environment packaging.** Converter-built `adapter-wheels-v1` packages include a complete wheel closure and do not require network access at job start. Conversion still requires an internet-connected host. Offline `wheels-v1` jobs likewise need a complete closure for the training image's Python and node architecture. * **DPO runtime.** Customizer `rl` DPO jobs run on Kubernetes/Ray and do not have a local Docker fallback. DPO remains full-weight only. * **Agent intent migration.** `AGENT-SPEC.md` is legacy. Use `ETHOS.md` for new work, and migrate a platform-owned legacy `AGENT-SPEC.md` package to `ETHOS.md` before relying on optimizer workflows. * **Optuna HPO runtime.** Optuna-based hyperparameter optimization runs locally from a developer workstation. It is not available as a horizontally scaled cluster service in this release. * **Access keys and workload delegation.** Scoped access keys and workload credential delegation require administrator configuration before workloads can rely on them in shared deployments. * **Embedded ClickHouse scope.** Helm's embedded ClickHouse is intended for development, evaluation, and non-critical single-node deployments. Use an externally managed ClickHouse for production deployments that require high availability, backups, or larger scale. ## Known issues * **Hermes trials can omit the `relay_atif` evidence descriptor.** Use the `trace:atif` descriptor instead; it points at the same ATIF file. * **DPO jobs fail before the first training step with `KeyError: 'checkpoint'`.** Customizer `rl` DPO jobs submitted from Studio, the CLI, or the API fail shortly after the training pod starts. The generated DPO configuration omits the `policy.dtensor_cfg.checkpoint` section that the pinned NeMo-RL version requires. There is no workaround in the 0.7 release line. DPO jobs run on v0.6.0. GRPO jobs are not affected. ## Links * Repository: [https://github.com/NVIDIA-NeMo/nemo-helix](https://github.com/NVIDIA-NeMo/nemo-helix) * Issues: [https://github.com/NVIDIA-NeMo/nemo-helix/issues](https://github.com/NVIDIA-NeMo/nemo-helix/issues) * Run the Agent Optimizer: [/documentation/agents/optimize-agents/run-the-agent-optimizer](/documentation/agents/optimize-agents/run-the-agent-optimizer) * Execute Agents as Jobs: [/documentation/agents/execute-agents-as-jobs](/documentation/agents/execute-agents-as-jobs) * Evaluate with a NeMo Fabric Harness: [/documentation/evaluate-models/agent-eval/fabric-runner](/documentation/evaluate-models/agent-eval/fabric-runner) * Using Authentication: [/documentation/access-control/authentication/using-authentication](/documentation/access-control/authentication/using-authentication) * Configure Sandboxed Gym for Evaluator: [/documentation/evaluate-models/agent-eval/gym-sandbox-configuration](/documentation/evaluate-models/agent-eval/gym-sandbox-configuration) * Install NeMo Helix Helm Chart: [/documentation/kubernetes-deployment/setup/helm/install](/documentation/kubernetes-deployment/setup/helm/install) * NeMo Fabric: [https://docs.nvidia.com/nemo/fabric/about-nemo-fabric/overview/](https://docs.nvidia.com/nemo/fabric/about-nemo-fabric/overview/) * NeMo Agent Toolkit: [https://docs.nvidia.com/nemo/agent-toolkit/latest/](https://docs.nvidia.com/nemo/agent-toolkit/latest/)