Overview

What a Request Package is

A RunLocal Request Package describes the computation you want to optimize: its operations, inputs, and what must stay the same.

Our agentic ML compiler uses it to understand what it can change. You can share this description without uploading your full source code or model weights.

The package, called a bundle in the API, contains a request file and its model files:

• Model structure: the operations and how they connect.
• Input details: shapes, data types, and example calls.
• Weight metadata: names, shapes, and data types. Your model weights stay local.

Review any included constants, sample data, or source code before remote validation or upload. They will be sent to RunLocal.

See complete packages in Model examples and field definitions in the Schema reference.

Your coding agent does the preparation

Choose the model or part of your stack and what you want to improve. With the RunLocal skill, your coding agent inspects your code, prepares the package, and asks for missing details.

Your agent can submit the package directly, or you can upload it in the workspace.

Review what will be sent before submission. Follow the Quickstart for setup and upload.

What happens after submission

Submission starts the assessment. You approve a target and quote before optimization starts.

1. Review targets and pricing. RunLocal proposes performance targets it estimates are achievable on your target hardware, with a fixed price and estimated timeline for each option.
2. Approve a target and quote. Once you agree, our agentic ML compiler runs the optimization on our infrastructure.
3. Download the results. Your RunLocal Deliverable Package includes optimized code or a binary, a performance report, and integration instructions.
4. Apply and verify. Have your coding agent apply the package to your onboard stack, then verify correctness and performance on your target hardware.

You only pay for an agreed target that is achieved. If no agreed target is met, you pay nothing.