RunLocal AI

RunLocal is an agentic ML compiler for edge compute platforms. It is built for teams who run autonomy and perception stacks on onboard hardware and want more performance from the same compute. Supported platforms today are Orin, Thor and Dragonwing, with more planned. The process works with a user's existing coding agent. The agent uses the RunLocal Skill to prepare a Request Package from the internal model stack locally, which the user reviews before the agent uploads it. The vendor states that no weights or training data are uploaded and that full source code is not required, and for open source models the repository and target hardware and software versions are enough. RunLocal's compiler then works autonomously on its own infrastructure, experimenting with custom kernels, quantization, operator fusion and scheduling. It does not currently perform distillation or broader model architecture changes. Optimized code, results and insights return as a Deliverable Package that the agent downloads and integrates, and the user validates the results locally. It is described as most useful where standard compilers and libraries fall short, such as dynamic workloads like sparse LiDAR perception, complex pipelines with several models sharing hardware, and unusual operations. A hardware reasoning engine predicts latency and remaining headroom for supported workloads, which helps teams judge what will fit and whether cheaper compute would meet their targets. Pricing is not published. The vendor invites prospective customers to book a call to discuss their stack and a possible one week pilot, and documentation is available.