DocumentationReferenceRelease Notes

v0.6.0 - 2026-09-23

View as Markdown

NeMo Platform v0.6.0 expands the 0.5 agent, evaluator, customization, and Studio workflows into a broader self-managed platform release. The release adds sandboxed NeMo Gym evaluation, retrieval evaluation support, source-checkout agent packaging, deployment-backed agent sessions, richer Studio workflows, workload credential delegation, and operational hardening for Kubernetes, Docker, and local development.

Highlights

  • Sandboxed task-driven evaluation. Agent evaluation can run NeMo Gym tasks in a sandboxed host, capture model calls, project Gym trajectories to OTLP, publish trial traces to Intake, and use clearer Harbor and Gym result models.
  • Retrieval evaluation and retrieval SDG. Evaluator adds BEIR and dense retrieval helpers, NIM embedding and ranking integrations, retrieval metrics, and value types. Data Designer adds retrieval synthetic-data generation for Nemotron models.
  • Fabric-backed agent lifecycle. Agents can be packaged from source checkouts, published as images, executed as jobs with inline agent definitions, invoked through deployment-backed sessions, and wired to Intake telemetry with less manual setup.
  • Studio workflow depth. Studio adds agent creation from uploaded directories, a shared Run Evaluation wizard, experiment trend and layout controls, and Customizer quick-start templates.
  • Auth and deployment hardening. Workload delegation and token exchange support credential propagation for deployed workloads. Kubernetes workloads gain default pod metadata and scheduling controls, and agent deployments fail earlier when their image or executor configuration is invalid.
  • Release-process automation. Fern release-version materialization can include release branches, stable GitHub releases are created by release automation, and forward-merge alerts now report failures to the on-call path.

What’s included

Agent Evaluation

  • NeMo Gym evaluations can run in a sandboxed Gym host instead of sharing the evaluator process. The sandboxed path supports a locked nmp-gym-host image, the nmp-cpu-tasks task image, broker placement controls, rollout streaming, and long-running rollout batches.
  • Gym trials can capture model calls and project ng_trajectory data into OTLP spans. Trials with traces can publish those traces to Intake so Evaluator, Experiments, and optimization workflows can inspect the same evidence.
  • The Gym runtime exposes more public cross-module names, clearer task and trial result values, typed runner reward keys, optional output aggregation, skill-use metrics, and better error rollups.
  • Harbor task suites use a stronger trial adapter and improved resume and reconciliation behavior, with better error propagation into the shared agent evaluation result model.
  • Retrieval evaluation support includes BEIR loading, dense search, NIM embeddings, NIM ranking, retrieval value objects, and retrieval metrics.
  • Evaluator documentation now covers Gym runners, custom Gym environments, sandboxed Gym configuration, Fabric harness evaluation, scoring by component, target and runner configuration, metric authoring, and result interpretation.

Agents and Fabric

  • Fabric agents can be packaged directly from a source checkout, including staged relative artifacts, packaged image publication to a registry, and simpler CLI flows for source packages.
  • Execute-agent jobs can accept inline agent definitions and have SDK support, CLI coverage, extension hooks, and documentation for batch execution through the platform jobs system.
  • Deployment-backed sessions support chat lifecycle operations so long-running deployed agents can keep session state behind the OpenAI-compatible gateway.
  • Agent telemetry can be wired automatically into Intake export paths, including trace attribution by agent name and richer logging for task execution, execute-agent jobs, and analyst runs.
  • Fabric adapter IDs can be fully qualified, and agent configs can bind multiple models.
  • Agent deployments now reject missing container images, unsupported deployment modes, and executor/backend mismatches before creating records. Kubernetes agent roots are staged into writable directories when mounted read-only.
  • Agent documentation adds the ETHOS.md optimization contract, agent environment guidance, and executing agents as jobs.

Studio

  • Studio can create Fabric agents by uploading an agent directory and has refreshed sample-agent flows.
  • Agent detail pages have cleaner evaluation tables.
  • Run Evaluation is now a wizard over a shared experiment form, with support for uploading agent-eval data and configuration from Studio.
  • Experiments gain metric trend charts, favorite markers, saved column layouts, and improved detail-page metrics.
  • Customizer adds quick-start templates, defaults from model specs, GRPO input and metrics fixes, clearer job failure banners, and improved job log loading.
  • Anonymizer is enabled by default in Studio feature flags for this release line.

Customizer and Data Designer

  • Customizer adds a reranking recipe and model-spec head type support.
  • Automodel training configuration and checkpoint handling are more complete, including embedding contract outputs and validation coverage.
  • GRPO configuration, documentation, Studio forms, hyperparameter handling, and serialization paths are hardened for RL and SFT workflows.
  • Data Designer adds retrieval synthetic-data generation for Nemotron models, plus CLI and provider updates for filesystem-backed seed handling.

Platform, CLI, and Deployment

  • Auth adds workload delegation, workload token exchange, and plugin SDK helpers for workload identity. Scoped access-key behavior is tightened in API and test coverage.
  • The CLI adds job log tailing, richer config help, TLS and context-delete support, improved setup behavior when NVIDIA Build probe models have reached end of life, and generated reference coverage for more plugin commands.
  • Kubernetes workloads can receive executor-level default pod annotations, selectors, tolerations, and affinity settings.
  • Model and inference paths improve Docker LoRA sidecar commands, deployment backed provider links to model entities, string /v1/models response handling, upstream inference-gateway error rewriting, and client x-api-key stripping before upstream proxying.
  • The platform runner reports controller thread failures as unhealthy, and ClickHouse readiness failures are surfaced more clearly.
  • Helm and Docker assets add OpenSandbox and Gym-host guidance, clearer chart and app-version placeholders, locked Gym-host dependencies, and release image maintenance.
  • Release automation now supports nightly wheels, stable GitHub Release creation for each release scope, and forward-merge alerting.

Install

For fresh local installs, see 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 Platform Helm Chart. Sandboxed Gym evaluation and GRPO reward environments also require OpenSandbox and job-storage configuration; see Configure Sandboxed Gym for Evaluator, GRPO and Reward Environments, and OpenSandbox.

Upgrade from v0.5.x

From an existing local checkout:

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:

nemo setup

After setup, restart local services before using CLI, SDK, Studio, or plugin workflows against the upgraded checkout.

If you run Kubernetes agent deployments, review your deployment executor and default image configuration before redeploying agents. v0.6.0 rejects missing images and deployment-mode mismatches earlier than previous releases.

If you use task-driven evaluation, update Gym and Harbor evaluation configs to the documented runner fields before moving production suites to sandboxed Gym execution.

Compatibility

  • Python 3.12-3.13
  • 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-platform Python SDK 0.6.0

Current constraints

  • Self-managed scope. v0.6.0 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. adapter-wheels-v1 packages require network access at job start for the built-in verifiers_agent path. Fully offline wheels-v1 jobs need a complete wheel 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 run nemo agents ethos migrate for platform-owned legacy packages 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.