AWQ-quantized Qwen2.5-7B-Instruct with layers 8-12 duplicated, ready to serve via vLLM. The math circuit runs twice on every forward pass.

28 base layers → 32 after duplication. No training, no merging, no weight changes — AWQ-quantized from the same RYS-modified weights as the GGUF companion.

The math specialist (vLLM-serving build). Same model and same sweep result as the GGUF companion, packaged for AWQ-quantized vLLM serving. The first sub-14B model the method was extended to — David Ng demonstrated layer-duplication on Qwen2-72B; we found it works just as cleanly here, 10× smaller. Of 51 swept configurations, (8,12) is the one that gets the math circuit alone.

vllm serve john-broadway/Qwen2.5-7B-RYS-8-12-AWQ \
    --quantization awq \
    --served-model-name Qwen/Qwen2.5-7B-RYS \
    --max-model-len 16384 --port 8000

51 configurations tested. Full sweep data in the v2 corpus dataset. For llama.cpp serving with GGUF Q4_K_M, see companion repo john-broadway/Qwen2.5-7B-RYS-8-12-GGUF.

Part of the RYS Sovereign Collection v1.

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

v2 — cross-architecture extension. 21 model variants across 10 architecture families. Headline: weak baselines lift more, in their weakest dimension. → john-broadway/rys-sovereign-collection-v2

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