I let my own 31B model take over development of the thing running it
Two things I build: Flash , a local coding agent that runs in your terminal. Flash Onyx , the model it runs on. Gemma4 with a system prompt and sampling baked in. As of today, I develop both of them from inside Flash, with Onyx driving. /model natuworkguy/flash-onyx-2.3:31b-cloudbase That tag is 96 KB on disk, because there are no weights in it. It is my prompt and my sampling pointed at a hosted…
In an unprecedented development, I entrusted my own 31B model, Flash Onyx, to take control of its own development process. The model, which runs in a terminal-based Flash agent with a system prompt and sampling baked in, operates on the Gemma4 architecture. The entire setup, weighing in at a mere 96 KB, relies solely on the system prompt and sampling to a hosted base, eliminating the need for hefty weights on the machine. The 12B variant, if desired, can run fully locally.
The core of the system is a system prompt, Onyx's obedience to which is documented in a modelfile within the repo. If the model's actions prove troublesome, I issue corrective instructions directly within the same session. This approach, coupled with the Python-based tools integrated within the model's tree, allows for seamless operation and iteration.
Unlike traditional evaluation suites that provide numerical outputs, the true value of this setup lies in its ability to expose subtle, systemic issues. These are not bugs, but subtle tics that can be detected only by human observation over prolonged periods. The most significant rule to emerge from this experiment is a self-policing mechanism that prompts the model to verify its actions before proceeding.
With a built-in safeguard against overstepping its bounds, the model is designed to request confirmation before committing to any changes. This ensures that the model remains accountable and does not quietly write itself a permission slip. The model's development branch, 2.4, incorporates these safeguards, with a new rule mandating evidence for every 'DONE' line in a report and demanding that shell calls be explicitly stated.
The project, accessible on Ollama and hosted on GitHub under an MIT license, is now live and ready for others to run and experiment with. With the model operating under a self-imposed, iterative governance model, I am keen to hear about any potential pitfalls or unexpected behaviors that others may encounter.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.