SmolLM2-135M-RYS-18-22-GGUF

SmolLM2-135M-RYS-18-22

SmolLM2-135M-Instruct with layers 18-21 duplicated. The late-stack reasoning + EQ circuit runs twice on every forward pass.

30 base layers → 34 after duplication. No training, no merging, no weight changes.

Reasoning 17.65% → 35.30% (+17.65). EQ 44.53 → 57.58 (+13.05). Math 0.315 → 0.303 (−1.20).

Results

The tiniest responder. SmolLM2-135M is the smallest model in the v2 corpus by an order of magnitude. RYS lifts both reasoning (+17.65 absolute) AND EQ (+13.05) simultaneously — the response is unremarkable on its own, but the comparison to sibling SmolLM2-1.7B (which lifts zero percent on reasoning) makes this card load-bearing: it falsifies the "SmolLM2 training recipe doesn't work with RYS" hypothesis. The 1.7B negative result is uniquely anomalous within the family, not architectural.

Pick this when you want the smallest possible model with reasoning + EQ lift. At 110MB Q4_K_M, this is the lightest RYS-applied checkpoint in the collection.

Usage

llama-server -m SmolLM2-135M-RYS-18-22-Q4_K_M.gguf -ngl 99

Full sweep data

40 configurations tested. (18,22) block-4 is the best-combined pick. Full per-config sweep + cross-architecture analysis: v2 dataset.

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. 13 deployable RYS-applied weight repos covering every non-zero-lift variant.

SmolLM2 family picture (all Q4_K_M):

• 135M (this card) — baseline reasoning 17.65%, peak Δ +17.65% (responds)
• SmolLM2-360M-RYS-12-15-GGUF — baseline reasoning 29.41%, peak Δ +23.53% (responds)
• SmolLM2-1.7B-RYS-eval — baseline reasoning 58.82%, peak Δ +0.00% (does NOT respond; eval-only — no RYS-applied weights since no lift to deliver)

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 on Qwen2-72B; sweep + probe toolkit by alainnothere.