Mycelium
Most software work sent to a frontier model can be done by local minions instead — given the right harness around them. Mycelium is a local AI workbench and the floor of a local Mac model factory: the models the agents run on are built, calibrated on the lab's own traces, measured, and served here, on hardware we own. We publish what the field rarely reports together — how much of the work runs locally, the energy each token costs, and the whole measurement ladder behind every model we ship, losing rungs included.
The lab, running
as of 2026-09-01 18:28 UTC · pushed from the lab's substrate at each deploy — failures included
Approach
A frontier model is reserved for judgment — decomposition, specifications, the calls that need taste. A roster of role-specialized local minions does the token-heavy execution. We publish what that division actually buys: the share of tokens carried locally, and the watts and joules each token costs on hardware we own.
This follows the intelligence-per-wattframing — accuracy per unit of power — and extends it from single-turn coverage to sustained, multi-step software work. The open question we are instrumenting, not yet answering, is whether the local share rises over time as the roster trains on the supervisor's own outputs, at held quality.
The factory closes the loop. The squad's own work traces calibrate the quantized models the squad then runs on — a measurable lever, not a slogan: two builds identical except the calibration corpus differ in behaviour, monotonically, at equal perplexity. Every build ships with its regime-stamped grid; the current production seat is published on Hugging Face with the losing rung beside it.