---

license: apache-2.0

base_model: Qwen/Qwen3.5-9B

language:

• en

tags:

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

---

RINIQ-MERNIK-GGUF — tri-fusion 9B quants

Finetune of Qwen/Qwen3.5-9B by

layer surgery (all donors descend from the same Qwen3.5-9B ancestor).

RINIQ is a new finetune built by layer surgery, quantized by the MERNIK

engine. Three Qwen3.5-9B donors, one 427-tensor trunk (Ornith's MTP head

excluded), per-block donor maps — no averaging of divergent blocks, ever.

Code and protocol: wepiqx/MERNIK

(full ledger in FUSION.md there).

Vision — natively multimodal family

Qwen3.5-9B is a natively multimodal family (early-fusion vision+language,

jointly trained — not a bolted-on encoder). This repo ships

mmproj-F16.gguf (stock Qwen3.5-9B vision projector, ~876 MB) next to the

text quants: load both for image/video input, omit --mmproj for

text-only.

llama-server -m RINIQ-M2-MERNIK-5100-Q38.gguf --mmproj mmproj-F16.gguf --port 28082 -ngl 99 -c 8192 --jinja

Note (MERNIK honesty): the HE/PPL verdicts below certify the text trunk

only. Vision quality on fused RINIQ blocks is UNVERIFIED — no vision

bench yet.

> ❤ If RINIQ runs on your hardware, consider sponsoring @wepiqx — consumer-GPU research, every scar public.

Donors

• Ox = OxCoder-9B (agentic code distillate, HE 90.24%) — base + globals
• Neo = NeoHorse-1-9B (thinking model) — donor of blk 31
• Orn = Ornith-1.5-9B-MTP trunk (GDN hybrid) — donor of mid/late blocks

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

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

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

| RINIQ-M2-MERNIK-5100-Q38 👑 | 5.0 GB | 7.6130 | 47.47% | 92.07% (151/164) | 87.8 | 8 — own-qwen38 lens: lost everything, won verdict |

| RINIQ-M2-MERNIK-5100 | 5.0 GB | 7.5819 | 50.00% | 91.46% (150/164) | 87.2 | 6 |

| RINIQ-M1-MERNIK-5100 | 5.0 GB | 7.6242 | 51.01% | 90.85% (149/164) | 85.4 | 7 |

| RINIQ-M3-MERNIK-5100 | 5.0 GB | 7.7111 | 50.00% | 86.59% (142/164) | 83.5 | 13 — soup buried |

| RINIQ-M4a-MERNIK-5100 | 5.0 GB | 7.5898 | 51.01% | 91.46% (150/164) | 87.2 | 4 — ties M2 + M1, best decisiveness |

| RINIQ-M4b-MERNIK-5100 | 5.0 GB | 7.5743 | 48.48% | 89.63% (147/164) | 84.1 | 4 — Orn-25 stronger than Ox-25 by 1.8pp |

| RINIQ-M4c-MERNIK-5100 | 5.0 GB | 7.5863 | 47.98% | 90.85% (149/164) | 85.4 | 6 — seam not poisonous |

| RINIQ-M2-SMAPE-6500 | 6.8 GB | 7.5730 | 48.48% | 89.63% (147/164) | 83.5 | 5 — wrong utility, scar |

| RINIQ-M2-MERNIK-6500 | 6.8 GB | 7.5263 | 48.48% | 91.46% (150/164) | 85.4 | 8 — ties 5100: ceiling |

| RINIQ-M5-MERNIK-5100 | 5.0 GB | 7.6587 | 52.02% | 91.46% (150/164) | 84.8 | 7 — kitchen sink |

| RINIQ-M6-MERNIK-5100 | 5.6 GB | 8.5137 | 46.97% | 83.54% (137/164) | 79.9 (−3.6) | 23 — BASE RULES |

| RINIQ-M7-MERNIK-5100 | 5.0 GB | 7.6033 | 49.49% | 88.41% (145/164) | 86.0 (−2.4) | 11 — first auto-fused |

| RINIQ-M2-MERNIK-5100-MIX | 5.0 GB | 7.6001 | 48.99% | 89.63% (147/164) | 85.4 (−3.6) | 6 — novel geometry |

| RINIQ-M2-RMSE-Q38-6500 | 6.8 GB | 7.5411 | 50.00% | 88.41% (145/164) | 82.9 (−4.9) | 6 — takes GPQA, loses HE by 3pp |

| Ox-MSE-6500 (parent) | 6.8 GB | 7.5125 | 49.5% | 90.24% (148/164) | — | 4–5 |

| Ox-SMAPE-5100 | 5.0 GB | 7.5670 | 49.0% | 88.41% (145/164) | — | 4–5 |

Recipes: M1 = Ox + Orn blks 15,19,23,27 (weight compass) + Neo blk 31.

M2 = Ox + Orn blks 24,25,26 (imatrix compass) + Neo blk 31.

M3 = M1 + blks 0–8 weight-averaged Ox+Neo (soup backbone) — soup buried

(−4.9pp vs M2, honest scar in FUSION.md).

M4a/b/c (finale slate, verdicts in): M4a = M2 + blk 16←Orn;

M4b = M2 with blk 25→Ox (one-block A/B on Orn-25);

M4c = M2 + blk 30←Orn (seam with Neo-31).

External reference: JackOD-9B-Coder (different harness!)

Corporate 4-way omnimerge_v2 over the same Qwen3.5-9B ancestor

(DeltaCoder 0.55 / Qwopus 0.30 / Ornith 0.15). Their numbers are quoted

from their README — greedy temp 0.0, mixed banks (their own † admits base

cells come from an earlier bank), NOT our protocol. Same-harness duel

(LiveCodeBench v6 55 hard, custom_evaluator on our generations) is queued

and will replace this table.

| Build | HE | HE+ | LCB v6 55 hard |

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

| RINIQ-M2-MERNIK-5100-Q38 (ours) | 92.07 | 87.8 | ⏳ duel queued |

| JackOD-9B-Coder (Q6_K, theirs) | 88.41 | 82.32 | 78.18 |

We lead on their reported HE/HE+; they lead where we haven't measured yet.

No conclusions until one harness scores both.

Status

Series CLOSED: crown Q38 (92.07), no 93%. Files land as decided. Quants land here as verdicts complete.

Naming: RINIQ-M<i>-MERNIK-<size>[-Q38].gguf.

No file is uploaded before its HE verdict unless marked UNVERIFIED.