---
license: gemma
base_model: google/gemma-2-9b-it
tags:
  - rys
  - layer-duplication
  - reasoning-circuits
  - emotional-intelligence
  - gguf
  - sovereign-collection-v2
---

# Gemma-2-9B-RYS-14-18

Gemma-2-9B-it with layers 14-17 duplicated. The **early-layer reasoning circuit** — at ~33% depth in a 42-layer stack — runs twice on every forward pass.

42 base layers → 46 after duplication. No training, no merging, no weight changes.

**Reasoning 58.82% → 82.35% (+23.53). Math 0.539 → 0.5576 (+1.86). EQ 94.06 → 92.85 (−1.21).**

## Results

| Metric | Baseline | RYS (14,18) | Delta |
|--------|----------|-------------|-------|
| Math | 0.539 | 0.5576 | +1.86 |
| EQ | 94.06 | 92.85 | −1.21 |
| Reasoning | 58.82% | 82.35% | +23.53 |

**The early-reasoner.** Gemma-2-9B has the **highest baseline EQ in the entire v2 corpus** (94.06) and the earliest reasoning peak position — layer 14 of 42 (~33% depth), significantly shallower than the late-layer peaks at ~60% depth seen in Llama-3.1-8B and Mistral-7B-v0.3. Duplicating the early reasoning block lifts reasoning +23.53% while EQ stays essentially stable (−1.21). The clean reasoning gain without EQ trade-off — unlike Granite-3.1-1B-A400M's MoE-routing trade-off — suggests Gemma's instruction-tuning recipe routes reasoning through earlier, more isolatable layers.

Pick this when you want a strong reasoning lift and the highest baseline conversational fluency in the collection.

## Usage

```
llama-server -m Gemma-2-9B-RYS-14-18-Q4_K_M.gguf -ngl 99
```

## Full sweep data

48 configurations tested. (14,18) block-4 is the best-combined pick. Full per-config sweep + cross-architecture analysis: [v2 dataset](https://huggingface.co/datasets/john-broadway/rys-sovereign-collection-v2).


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. Gemma-2-9B sits slightly above the curve at its baseline — the shallow circuit position and clean EQ profile make it a notable distinct response within the corpus. 13 deployable RYS-applied weight repos covering every non-zero-lift variant.

**Within-family sibling:** [`john-broadway/Gemma-2-2B-RYS-19-23-GGUF`](https://huggingface.co/john-broadway/Gemma-2-2B-RYS-19-23-GGUF) — the matched-baseline smaller sibling that lifts less, demonstrating depth-room scaling.

**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](https://dnhkng.github.io/posts/rys/) on Qwen2-72B; sweep + probe toolkit by [alainnothere](https://github.com/alainnothere/llm-circuit-finder).