RunLocal

RunLocal is an agentic model compiler for edge compute platforms such as NVIDIA Jetson and Qualcomm. ML teams prepare a request package with their coding agent, RunLocal optimizes it on its infrastructure while project data, weights and training data stay local, and the team downloads a deliverable package to apply and validate the optimizations.

The Request Package includes model structure, operations and shapes, hardware and software settings, representative test inputs, the objective, and optional inline operator source. The Deliverable Package contains optimized code or a binary, a performance report with measured results and estimated remaining headroom, and integration instructions for the coding agent. The platform uses custom MLIR dialects with a device simulator and performance modeling to attribute estimated performance costs to implementation choices, plus on-device orchestration that schedules experiments, reserves hardware and cleans up runs. A Torch request can include a small Python callable defining one graph operator.

RunLocal is aimed at ML teams in physical AI, including robotics, drones, autonomous vehicles and industrial perception. Teams set performance targets, agree to a fixed fee, and pay only if the target is achieved. For open-source models, the setup steps are not required.