---

license: apache-2.0

base_model:

• XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B
• OrionLLM/OxCoder-9B
• ornith-ai/Ornith-1.5-9B
• TokenRhythm/NeoHorse-1-9B
• Qwen/Qwen3.5-9B

language:

• en

tags:

• quantization
• gguf
• llama-cpp
• fusion
• riniq
• mernik
• humaneval

---

RINIQ-NEXT-GGUF — Ox × MiMo layer fusions

RINIQ-NEXT continues the RINIQ program on a new donor pair: OxCoder-9B × MiMo-V2.6-Distill-Qwen-9B (same Qwen3.5-9B bones, wildly different finetunes — agentic coder vs agentic SFT distillate, 77B SFT tokens). Layer surgery via fuse_layers.py, quantized by the MERNIK engine (SMAPE-5100, Ox+Neo dual-max lens — the exact M2 recipe, for clean A/B). Method + full ledger: wepiqx/MERNIK (RINIQ-NEXT.md, MIMO.md, FUSION.md).

> ❤ Consumer-GPU research — consider sponsoring @wepiqx. Every scar public.

Status: lab bench (no weight uploads yet)

Honest call: nothing here dethrones regular RINIQ on code (M2-Q38 92.07%, M2-5100 91.46% / HE+ 87.2). N2 ties the 5100 crown but loses on rigor (HE+ 85.4 vs 87.2) — a tie is not a reason to ship 5 GB. Mutants stay local until something beats M2; this repo tracks verdicts, files land only with a crown.

Recipes

• N1: Ox/MiMo alternating interleave + MiMo blks 15,16,17 (built before the --map fix — misread as a graft, scar kept).
• N1m: mirror interleave + Ox blks 15,16,17.
• M8 (single-swap A/B): RINIQ-M2 base (Ox + Orn 24,25,26 + Neo31 baked in) + MiMo blk 31. Edge position = minimal seam surface.
• N2 (mashup 🙏): RINIQ-M2 base + MiMo blks 15,16,17. Donor islands (15–17 vs 24–26) far apart — hypothesis: distant islands interfere less.

Building from RINIQ-M2-BF16 as base replaces three donors at once (Neo BF16 isn't local — Neo31 travels verbatim inside M2).

Verdicts (slow ring: HumanEval pass@1, temp 1.0 / top_p 0.95 / top_k 20, presence 0.0, max 2048)

| Build | Size | PPL (ctx1024) | HE pass@1 | HE+ | Empties |

|:------|-----:|:-------------:|:---------:|:---:|:-------:|

| RINIQ-N2 (M2+MiMo15–17) 🙏 | 5.0 GB | 7.6813 | 91.46% (150/164) | 85.4 (−6.1) | 8 — TIES M2 crown, distant islands confirmed |

| RINIQ-M8 (M2+MiMo31) | 5.0 GB | 7.6479 | 82.93% (136/164) | 79.3 (−3.6) | 15 — MiMo31 loses to Neo31 by 8.5pp, clean negative |

| RINIQ-N3 (M2+MiMo10,14,15,16,27 rank-vote) | 5.0 GB | 7.6313 | 88.41% (145/164) | 82.9 (−5.5) | 9 — machine goes 0-2 (M7 scored exactly the same 88.41%) |

| RINIQ-N4a (M2+MiMo15–17 FFN-only) 👑 | 5.0 GB | 7.6276 | 91.46% (150/164) | 86.6 (−4.9) | 5 — soul is in FFN: tissue alone ties the crown, tidiest code |

| RINIQ-N4b (M2+MiMo15–17 attn-only) | 5.0 GB | 7.6163 | 89.02% (146/164) | 84.1 (−4.9) | 12 — soul divisible, FFN-dominant |

| RINIQ-N4c (M2+MiMo15–17 norms-only) | 5.0 GB | 7.5795 | 89.63% (147/164) | 85.4 (−4.2) | 7 — best canary, sub-crown verdict (mild liar) |

| RINIQ-N4a-Q38-SMSE (soul+lens+blend) | 5.0 GB | 7.6617 | 89.02% (146/164) | 84.8 (−4.2) | 14 — STACKING LOSES: three separate winners interfere, wins don't sum |

| RINIQ-N1 (interleave) | 5.0 GB | 8.0572 | 79.27% (130/164) | 76.2 | 24 |

| RINIQ-N1m (interleave) | 5.0 GB | 7.9604 | 63.41% (104/164) | 63.4 (−0.0) | 51 — best PPL, worst verdict: PPL-liar crown 👑 |

| RINIQ-M2-MERNIK-5100 (base for M8/N2) | 5.0 GB | 7.5819 | 91.46% (150/164) | 87.2 (−3.0) | 6 |

| Ox-SMAPE-5100 (Ox family) | 5.0 GB | 7.5670 | 88.41% (145/164) | — | 4–5 |

| Ox-Q6_K (stock, Ox family) | 7.2 GB | 7.6758 | 84.15% (138/164) | — | — |

| MiMo-5100-SMAPE (MiMo family) | 5.0 GB | 8.7435 | 70.12% (115/164) | 66.5 | 7 |

| MiMo-Q6_K (stock, MiMo family) | 7.2 GB | 9.1362 | 71.95% (118/164) | 66.5 | 6 |

Reading the table: one block decides (blk31 Neo→MiMo costs 8.5pp — the V2→M2 law in reverse), distant donor islands integrate free (N2 ties M2 exactly), and perplexity can crown the worst model (N1m: best PPL 7.96, HE 63.41%, 51 empties — the strongest canary-vs-verdict split on record). Boldness tax stands: N2's rigor drop (−6.1) vs M2's (−3.0).

Tissue mode (fuse_layers.py: 15:b:ffn, ranges 15-17:b:ffn, multi ffn+attn, ln for any norm tensor, --dry, --d): N4a proves the soul is in FFN (tissue alone ties the crown at 91.46% with 5 empties); N4b (−2.4pp) shows attention carries compatibility; N4c sets the trap with the best PPL of all.

Usage

llama-server -m RINIQ-N2-MERNIK-5100.gguf --port 28082 -ngl 99 -c 8192 --jinja