SmolLM2-360M-RYS-12-15-GGUF

SmolLM2-360M-RYS-12-15

SmolLM2-360M-Instruct with layers 12-14 duplicated. A mid-stack reasoning + math circuit runs twice on every forward pass.

32 base layers → 35 after duplication. No training, no merging, no weight changes.

Reasoning 29.41% → 52.94% (+23.53). Math 0.279 → 0.3688 (+8.98). EQ 77.66 → 72.74 (−4.92).

Results

The middle SmolLM2 sibling. Swept with the full block-size search [3,4,5] across 66 configurations — the same search depth as the rest of the v2 corpus, broader than the SmolLM2-1.7B sibling's narrower [3,4] search. Responds normally with +23.53 reasoning and a +8.98 math gain, reinforcing the 135M's signal: the 1.7B negative result is the anomaly, not the family. The narrower 1.7B sweep remains a methodological caveat worth a re-sweep with full blocks to resolve definitively.

Pick this when you want a tiny model with strong reasoning + math lift and can absorb the small EQ trade-off.

Usage

llama-server -m SmolLM2-360M-RYS-12-15-Q4_K_M.gguf -ngl 99

Full sweep data

66 configurations tested. (12,15) block-3 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):

• SmolLM2-135M-RYS-18-22-GGUF — baseline reasoning 17.65%, peak Δ +17.65% (responds)
• 360M (this card) — baseline reasoning 29.41%, peak Δ +23.53% (responds; full block search [3,4,5])
• 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.