rndhouse/mixmod

Mixmod is a Rust project that tests whether an expensive, high-capability AI model can supervise a cheaper local model to do software engineering work, while preserving the quality of results. The specific pairing tested is GPT-5.5 acting as the supervisor and Qwen3.6-27B as the worker model, running on a single RTX 3090 graphics card. The core idea is a loop. GPT-5.5 gives instructions to the local Qwen model, which then modifies code. GPT-5.5 reviews the changes. If the work is approved, the task is complete. If not, GPT-5.5 prepares feedback and sends it back to the worker model for another attempt. This structure lets the expensive supervisor model spend fewer tokens because the cheaper worker model handles the actual code editing. Early benchmark results on a 10-task pool from SWE-bench Lite show a 75.5% aggregate reduction in GPT-5.5 token usage. Output tokens from the supervisor fell by 51.4%, and input tokens fell by 76.1%. Per-task total token reductions ranged from 56.0% to 91.4%. The project is careful to note that this benchmark pool is selected from tasks GPT-5.5 could already solve on its own. Mixmod is testing token reduction, not trying to improve the raw capability of the models. The README includes a per-task breakdown showing how much each benchmark task reduced GPT-5.5 input and output tokens. Reductions vary across tasks. For example, one Django task cut input tokens by 91.9% and output tokens by 73.1%, while another Django task cut input by 56.3% and output by 22.0%. The project points readers to a separate benchmark document for methodology, runtime details, and caveats. For installation, the README lists Codex and OpenCode as requirements. You install Mixmod via a single command using Cargo, the Rust package manager, and then run it by passing a task description in quotes. The README is short and does not provide further setup guidance, troubleshooting, or configuration details beyond this quick start example.