---

license: apache-2.0

language:

• en

base_model: OrionLLM/OxCoder-9B

pipeline_tag: text-generation

library_name: gguf

tags:

• quantization
• gguf
• llama-cpp
• imatrix
• hybrid-quantization
• MERNIK
• humaneval
• coding

---

OxCoder-9B — MERNIK Quantization

MERNIK ("the one who measures") quants of

OrionLLM/OxCoder-9B (Qwen3.5-9B agentic

coding finetune, 262K context): imatrix-driven priority-queue allocation,

judged by **all three columns in synergy — PPL (canary), KLD-vs-ref (rank),

scored tasks (verdict)**.

Results (all builds, all columns)

HE = HumanEval, 164 tasks, pass@1. llama-server --jinja, temp 1.0 /

top_p 0.95 / top_k 20 / presence 0.0, max_tokens 2048, fixed seeds.

PPL/KLD: llama-perplexity -ngl 99 -c 1024 -n 64 -b 512 --seed 7,

wikitext-2-raw.

| Build | Size | PPL | KLD vs Q8 | HE pass@1 |

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

| OxCoder-9B-MERNIK-6500-Q8.gguf | 6.83 GB | 7.5125 | 0.0386 | 90.24% (148/164) 👑 — beats NeoHorse-MSE by +4.3pp at same weight |

| OxCoder-9B-MERNIK-6500-SMSE.gguf | 6.83 GB | 7.5410 | — | 91.46% (150/164), HE+ 87.8 — thank-you build TAKES the Q8 crown by +1.2pp |

| OxCoder-9B-MERNIK-5100-Q4.gguf | 5.36 GB | 7.5670 | 0.0505 | 88.41% (145/164) — small beats NeoHorse big: distillate depth |

| Q6_K (stock) | 7.36 GB | 7.6758 | 0.0116 | 84.15% (138/164) — flat keeps up, allocation still wins |

Finetune-vs-finetune duel: distillate leads on all weights.

> ⚡ Fast ring: GPQA-recognition (non-standard). Ox-MSE-6500 49.5%,

> Ox-SMAPE-5100 49.0% — below Neo despite winning HE: thinkers spread

> first-token mass. Recognition ≠ reasoning; details in

> wepiqx/MERNIK.

Method + protocol: wepiqx/MERNIK.