README.md

metadata

title: SoftBacon Software
emoji: 🥓
colorFrom: red
colorTo: gray
sdk: static
pinned: false

SoftBacon Software

A small independent lab working on locally-owned AI: models that run on hardware you own, measured honestly, with the measurement published alongside the weights.

Everything here is produced on three machines, by one person and a squad of local agents that live on them:

• an Apple M5 Max — where the models we build and serve actually run;
• a single RTX 3090 — the second inference box, reached over a private network;
• a Jetson Orin Nano (8 GB, ARM) — always on, running the coordination substrate the agents share: their memory, their tasks, their messages to each other. It serves no model. It is the part that makes them a crew rather than a set of processes.

No rented GPUs are in that list. Anything trained on borrowed hardware will say so on its own card.

What we publish

Quantizations with a measured ladder. Not a single "best" build, but the whole series with its losses and its speeds, including the rungs that lost. A ladder with the bad rungs removed is marketing, not measurement.

Adapters, with the corpus and the gate that admitted them.

The harness, so a row in our tables can be reproduced rather than trusted. Build scripts and benchmark invocations live at github.com/SoftBacon-Software.

How we report numbers

• Accuracy on standard suites — HumanEval / HumanEval+, Aider-Polyglot, held-out loss — never an internal metric quoted alone.
• Speed on two separate axes, because they answer different questions: plain autoregressive decode, and the speculative path.
• Every table carries its regime stamp: the harness commit, the server settings, and n. A number measured under a different regime is a different number, and we will not put the two in one table.
• Negative results are published. A build that loses on both axes is a row, not a deletion.

Status

First release, 2026-08-24: a controlled pair of 4-bit MLX quantizations of Laguna XS 2.1 — Laguna-XS-2.1-sbQ-cal-4bit (imatrix calibrated on our own agent-squad traces; the build our lab serves) and Laguna-XS-2.1-sbQ-4bit (its generic-calibration control). Same tool, same commit, same bits — the comparison is the finding: calibration shifts behaviour and speed at equal perplexity. The full measurement regime and the field note are at mycelium.fyi.

sbQ marks SoftBacon builds — unrelated to poolside releases and to community oQ-series quants. Further rungs appear when they have been measured, and not before.