Llama-3.1-8B-RYS-18-22-GGUF

Llama-3.1-8B-RYS-18-22

Llama-3.1-8B-Instruct with layers 18-21 duplicated. The mid-late stack reasoning circuit runs twice on every forward pass.

32 base layers → 36 after duplication. No training, no merging, no weight changes.

Reasoning 82.35% → 94.11% (+11.76). Math 0.525 → 0.5885 (+6.35). EQ 87.38 → 81.79 (−5.59).

Results

The high-baseline multi-circuit lift. Llama-3.1-8B already has strong reasoning (82.35% baseline). RYS still finds 15 of 66 configurations that boost reasoning >5% — at comparable scale, Qwen2.5-7B-Instruct shows only 5 boosting configurations. Llama-3.1's deep stack carries multiple parallel reasoning paths; duplicating any one of them gives a measurable lift without breaking the others.

Pick this when you want stronger reasoning out of a well-trained 8B class model and can afford the modest EQ trade-off.

Usage

llama-server -m Llama-3.1-8B-RYS-18-22-Q4_K_M.gguf -ngl 99

Full sweep data

66 configurations tested. (18,22) block-4 is the best-combined pick. Full per-config sweep + cross-architecture analysis: v2 dataset.

Part of the RYS Sovereign Collection v2.

---

Where this sits in the Sovereign Collection

v1 — Qwen2.5 cross-scale + Qwen3-32B headline crossover. 5 model repos.

v2 — cross-architecture corpus. 21 model variants across 10 architecture families. Inverse correlation (r = −0.726): weak baselines lift more, in their weakest dimension. Llama-3.1-8B sits at the high-baseline end of the curve, where the richness of the multi-circuit signal (number of boosting configurations) is the load-bearing finding rather than per-config magnitude. 13 deployable RYS-applied weight repos covering every non-zero-lift variant.

Cross-architecture comparator: john-broadway/Mistral-7B-v0.3-RYS-18-23-GGUF — same circuit-position (layers ~18-22), 40-point weaker baseline (41.18% vs 82.35%), 28 boosting configurations vs Llama's 15.

Credit

John Broadway, with collaboration from Claude (Opus 4.6 in April 2026 sweep generation and build pipeline; Opus 4.7 in May 2026 cross-architecture analysis and publication). Original RYS method by David Ng on Qwen2-72B; sweep + probe toolkit by alainnothere.