Qwen3-8B-RYS-16-19-GGUF

Qwen3-8B-RYS-16-19

RYS-enhanced Qwen3-8B with layers 16-19 duplicated. 36 layers expanded to 39. Zero training, zero weight changes.

Math +6.7%, Reasoning +23.5%. Baseline reasoning healed from 53% to 76%.

Results

117 configurations tested. The 8B's baseline reasoning was the weakest in the Qwen3 family (53%). RYS at (16,19) heals it to 76%.

Usage

llama-server -m Qwen3-8B-RYS-16-19-Q4_K_M.gguf -ngl 99

Full sweep data

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

Part of the v2 Qwen3-family cohort — parallel Qwen3-family RYS-applied weights from April 2026, expanded alongside the v1 Qwen2.5 cross-scale collection. (The "four model scales" originally referenced here was a Qwen3-only expansion; the original v1 writeup described Qwen2.5 cross-scale + Qwen3-32B as headline.)

Where this sits in the Sovereign Collection

v1 — Qwen2.5 cross-scale + Qwen3-32B headline crossover (the original v1 intent per the 2026-04-11 writeup). 5 model repos on HuggingFace; see john-broadway.

v2 Qwen3-family cohort (this card's cohort — parallel Qwen3-family RYS-applied weights, April 2026):

• 0.6B — john-broadway/Qwen3-0.6B-RYS-10-13-GGUF
• 1.7B — john-broadway/Qwen3-1.7B-RYS-7-10-GGUF
• 8B (this card) — john-broadway/Qwen3-8B-RYS-16-19-GGUF

v2 cross-architecture corpus (21 model variants spanning 10 architecture families): john-broadway/rys-sovereign-collection-v2

Attribution: John Broadway, with collaboration from Claude (Opus 4.6 in April 2026 build; Opus 4.7 in May 2026 cross-architecture analysis and family-relabeling). 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 52.94%, peak RYS Δ +29.41%. Of 117 swept configurations, 71 boost reasoning >5% — the highest hit rate in the corpus. The (16,19) block is one of the consistent boosters at L16-19.

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.