Qwen3-32B-RYS-20-28-GGUF

Qwen3-32B-RYS-20-28

RYS-enhanced Qwen3-32B with layers 20-28 duplicated. 64 layers expanded to 72. Zero training, zero weight changes.

Math +4.5%, EQ +0.04, Reasoning +18%. All three metrics positive.

Results

The Big Boy. Baseline logic score was 0% — RYS fixes it. The (20,28) config is the only one that improves all three metrics simultaneously.

Usage

llama-server -m Qwen3-32B-RYS-20-28-Q4_K_M.gguf -ngl 99

Full sweep data

63 configurations tested. Sweep results published with the model files.

Part of the RYS Sovereign Collection v1 — the headline crossover.

Where this sits in the Sovereign Collection

v1 — Qwen2.5 cross-scale + Qwen3-32B headline crossover. Four sizes from 0.5B to 32B; RYS works at every scale, with the lift size and dimension shifting by baseline:

• 0.5B → EQ specialist (Qwen2.5-0.5B-RYS-3-7-GGUF)
• 1.5B → balanced daily driver (Qwen2.5-1.5B-RYS-4-7-GGUF)
• 7B → math specialist via (8,12) (Qwen2.5-7B-RYS-8-12-GGUF + -AWQ)
• 32B (this card) — the headline "Big Boy"

v1 attribution: John Broadway, with collaboration from Claude (Opus 4.6 in April 2026 build; Opus 4.7 in May 2026 v1 republication aligning to the original Qwen2.5 cross-scale intent). Original RYS method by David Ng on Qwen2-72B; sweep toolkit by alainnothere.

---

v2 cross-architecture context (2026-05-13)

This model's place in the v2 curve: baseline reasoning 70.59%, peak RYS Δ +17.65%. The (20,28) 8-layer block sits at 30–50% depth, the family-invariant band for deeper models. 38 boosting configurations make this the most-validated single-model dataset in v2.

Across the 21 model variants (10 architecture families) surveyed in john-broadway/rys-sovereign-collection-v2:

• Pearson r(baseline reasoning, peak RYS lift) = −0.726. Weak baselines lift more, in their weakest dimension.
• Three RYS-recoverable suppression mechanisms identified: under-training scale, MoE routing inefficiency, specialization training trade-off.
• One published negative result (SmolLM2-1.7B). RYS is not universal.

v2 attribution: John Broadway, with cross-architecture analysis by Claude (Opus 4.7). Original RYS method by David Ng; circuit-finder toolkit by alainnothere.