Nvidia

Open source activity

As of Oct 28, 2026

Nvidia's open-source activity splits cleanly into two lines right now. One is the long-running RAPIDS/CUDA software stack — cuDF (GPU dataframes), cuML (GPU-accelerated machine learning), and the CUDA Core Compute Libraries — all getting steady maintenance as the foundation under everything else Nvidia ships. The other, newer line is agent infrastructure: Model-Optimizer (a unified library for quantization, distillation, pruning and speculative decoding) and NemoClaw, which lets agent frameworks like Hermes, LangChain Deep Agents and OpenClaw run "more securely inside NVIDIA OpenShell with managed inference."

That NemoClaw repo is a direct, concrete piece of the containment story machine.net has already covered: OpenShell is the hardware-level kill-switch architecture Nvidia built specifically to let operators cut power to a misbehaving agent's compute, and NemoClaw is the software layer that routes third-party agent frameworks through it rather than letting them run with unmanaged access to raw GPU inference. Seeing active development on the integration layer, not just the announcement, suggests Nvidia is treating agent containment as a product surface to build out, not a one-time press release.

Most active repositories