NVIDIA has open-sourced OSMO. The Kubernetes-native workflow orchestrator targets physical AI pipelines that span data-center GPUs, workstation hardware, and edge devices without forcing teams to write infrastructure code for each step.
OSMO lets a team describe the whole pipeline in a single YAML file, then run that workflow across every computing tier. The project treats training, simulation, deployment, and hardware-in-the-loop testing as parts of one system — a useful change from stitching together separate tools and hoping the handoffs behave.
NVIDIA announced the release on September 14, 2026. The latest release is version 6.3.1, which arrived in June 2026, giving the project a defined release history before its public open-source push.
One control plane for three computers
NVIDIA frames physical AI as a three-computer problem. Training runs on data-center GPUs, simulation and sensor rendering use workstation-class RTX hardware, and deployment or hardware-in-the-loop testing runs on edge devices such as Jetson AGX Thor.
OSMO treats those environments as backends of one control plane. Each backend is a Kubernetes cluster registered through the OSMO command-line interface, so the workflow can move between computing tiers without changing the infrastructure layer inside the pipeline.
The canonical example uses three tasks: a simulation on rtx-pro-6000, training on gb200 with 8 GPUs, and evaluation on jetson-agx-thor. Testing environments also reference Isaac Sim, tying the workflow to simulation before models reach physical hardware.
That arrangement matters because physical AI development does not end when training finishes. A model must survive simulated sensors, deployment constraints, and tests on the target device — three environments that rarely share the same hardware or scheduling rules.
Open-source packaging with enterprise controls
OSMO carries an Apache-2.0 license and ships Helm charts and containers on NGC. A local quickstart runs the full control plane on a workstation with KIND, giving teams a way to test the system without first assembling a fleet of clusters.
Its feature set covers portability, interactive development, scheduling, data management, security and identity, and agent integration. The default scheduler is the NVIDIA KAI Scheduler, while the workflow model keeps the task definition separate from the infrastructure needed to execute it.
Data management is one of OSMO’s more concrete claims. The project supports content-addressable datasets with deduplication and says the approach can cut storage needs by 10 to 100x. Storage systems have heard promises like this before, but attaching the claim to a workflow orchestrator makes dataset reuse part of the pipeline rather than an afterthought.
The security layer includes an RBAC authorization sidecar, OAuth2 proxy integration, and TLS termination at the Envoy gateway. OSMO has also supported cloud workload identity since version 6.2.8, extending the control plane’s identity model beyond local cluster access.
The repository includes an AGENTS.md file, a skills directory, and an MCP deployment guide. At GTC 2026, NVIDIA said OSMO integrates with Claude Code, OpenAI Codex, and Cursor so coding agents can submit, monitor, and debug pipelines — because apparently the workflow now needs assistants watching the workflow.
The project information also lists 150k+ ML SubReddit members, underscoring the size of the machine-learning audience surrounding tools like this. OSMO’s real test will be whether teams adopt the single-workflow model when their training clusters, simulation workstations, and edge hardware refuse to behave like one tidy computer.
