---
license: apache-2.0
base_model: ibm-granite/granite-3.1-1b-a400m-instruct
tags:
  - rys
  - layer-duplication
  - reasoning-circuits
  - gguf
  - sovereign-collection-v2
  - mixture-of-experts
---

# Granite-3.1-1B-A400M-RYS-12-15

Granite-3.1-1B-A400M-Instruct (MoE, 400M active per token) with layers 12-14 duplicated. Mid-stack expert chain runs twice on every forward pass.

24 base layers → 27 after duplication. No training, no merging, no weight changes.

**Reasoning 23.53% → 76.47% (+52.94). EQ 84.41 → 70.89 (−13.52). Math 0.394 → 0.4707 (+7.67).**

## Results

| Metric | Baseline | RYS (12,15) | Delta |
|--------|----------|-------------|-------|
| Math | 0.394 | 0.4707 | +7.67 |
| EQ | 84.41 | 70.89 | −13.52 |
| Reasoning | 23.53% | 76.47% | +52.94 |

**The MoE reasoning unlock.** First MoE in the v2 corpus. Duplicating layers 12-14 lifts reasoning by +52.94 — but EQ degrades on every one of the 32 swept configurations (range −6.76 to −47.07). Hypothesis: layer duplication preserves a localized reasoning circuit (a tight expert chain) but disrupts EQ-relevant experts that route across more diverse paths. The within-family dense comparison ([`Granite-3.1-2B-RYS-19-24`](https://huggingface.co/john-broadway/Granite-3.1-2B-RYS-19-24-GGUF), 40 layers) shows weaker reasoning lift (+11.76%) and stable EQ — consistent with the MoE-specific hypothesis.

31 of 32 swept configs boost reasoning >5%. Pick this when reasoning gain matters more than the EQ trade-off.

## Usage

```
llama-server -m Granite-3.1-1B-A400M-RYS-12-15-Q4_K_M.gguf -ngl 99
```

## Full sweep data

32 configurations tested. (12,15) block-3 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: 0.5B EQ specialist / 1.5B daily driver / 7B math specialist (+ AWQ) / Qwen3-32B "Big Boy."

**v2 — cross-architecture corpus.** 21 model variants across 10 architecture families. Inverse correlation (r = −0.726): weak baselines lift more, in their weakest dimension. Three mechanisms identified: under-training scale (Llama-3.2-1B), MoE routing inefficiency (this model), specialization training trade-off (Qwen2.5-Coder-1.5B). Plus EQ-amplifier extreme (TinyLlama-1.1B) and a first published negative result (SmolLM2-1.7B). 13 deployable RYS-applied weight repos covering every non-zero-lift variant.

**Within-family sibling:** [`john-broadway/Granite-3.1-2B-RYS-19-24-GGUF`](https://huggingface.co/john-broadway/Granite-3.1-2B-RYS-19-24-GGUF) — the dense 2B sibling. The MoE-vs-dense contrast within the Granite family is one of the cleanest mechanism-isolating cases in the corpus.

**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).