MUTHALLATH-1 m0: an honest experiment log
MUTHALLATH ("triangle") is an architecture experiment. It adds geometric "triangle" organs to a small model (base Qwen/Qwen3-0.6B, Apache-2.0) to test whether structured geometry can help a tiny model do agent tool tasks that normally need far larger models.
No weights are published. The current result is negative.
Clean results (contamination-checked arena, 120 tool tasks)
| Contestant | Accuracy | 95% CI |
|---|---|---|
| muse-glimmer-30b (reference, 30B) | 98.33% | 94.1–99.5 |
| jabarr, triangles OFF (0.6B) | 13.33% | 8.4–20.6 |
| jabarr, triangles ON (0.6B) | 10.83% | 6.4–17.7 |
Correction: an earlier claim that "Jabarr ties Muse-30B at 96.7%" was a scoring artifact. The regex scorer gave credit to content-free output. The clean scorer (arena_clean.py, included here) refuses to score any item that leaked into the training corpora and prints a regex baseline next to every model score. On it, the triangles give no measurable gain yet (ON 10.8% vs OFF 13.3%, overlapping CIs).
Vision
The goal is a small model that can be audited and that reasons through explicit geometric and Arabic-root structure instead of brute scale. This log stays public so the failures count as much as the wins.
About the lab
ABUZ8 is a one-person lab. Ahmad builds it alone, with no team and no funding, and has spent the last year living inside AI work. The long-term goal is the SUPERBRAIN fusion: Arabic-root reasoning (SUPERBRAIN), tool use and memory (Qadir), fast local inference (FAIE) and new architecture experiments (MUTHALLATH), all joined in one open system.