MiMo-V2.6-Distill-Qwen-9B-GGUF-MERNIK
MERNIK quants of XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B — Xiaomi's agentic SFT distillate of Qwen3.5-9B (77B SFT tokens, 27% visual). Quantized by the MERNIK engine: instead of cutting every layer equally (like flat Q4_K_M / Q6_K), an importance queue spends each megabyte where it matters most. Own MiMo imatrix + qwen38 calibration, never mixed with foreign lenses. Method + ledgers: wepiqx/MERNIK.
❤ Consumer-GPU research — consider sponsoring @wepiqx. Every scar public.
Status: all builds landed
All six MERNIK builds + two stock references carry verdicts (table below).
Naming: MiMo-V2.6-Distill-Qwen-9B-MERNIK-<size>[-SUFFIX].gguf
(*-SMSE.gguf for the blend; the first two shipped as *-BALANCE.gguf,
same formula, old name). Stock Q3_K_M row: PPL + size measured, HE
battery running — verdict lands as measured.
The three lenses (plain words)
- SMAPE (barbell) — junk layers drop to the Q4 floor, kings rise to Q8 penthouses, almost nothing in the middle. Owns tight budgets.
- MSE (spread) — fills the middle (Q5/Q6), almost no penthouses. Owns big budgets.
- SMSE (new) — geometric blend of the two: a fat middle (Q5/Q6) with a smaller sub-4 dungeon. Kings without people are nobody, people without kings are nobody. Built because SMAPE's relative math silently cancelled the sub-4 toxicity penalty — the blend carries real toxicity through its absolute half.
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+ |
|---|---|---|---|---|
| MERNIK-5100-SMAPE | 5.0 GB | 8.7435 | 70.12% (115/164) | 66.5% |
| MERNIK-5100-SMSE | 5.0 GB | 8.7389 | 71.34% (117/164) | 65.9 — duel 0/3 (HE p=0.88, HE+ against, empties p=0.13): +1.2pp is NOISE, crown withdrawn 2026-09-28 |
| MERNIK-5100-SRMSE | 5.0 GB | 8.7389 | 70.12% / 66.46% (reruns) | ~64 — bit-identical file to SMSE-5100 (md5 match!), scores 71.34/66.46/70.12 across runs: pure sampling spread, no OC harm found |
| MERNIK-4500-SMAPE (allow-q3) | 4.4 GB | 10.5343 | 29.88% (49/164) | 26.2 — floor-is-lava zone (cf. NeoHorse-3650-Q2 23.78%): sub-4 at ~4GB, same physics |
| MERNIK-4500-SMSE (allow-q3) | 4.4 GB | 9.4092 | 54.88% (90/164) | 53.0 — rescue: +25pp over SMAPE, TOX-like save from the lava |
| MERNIK-6500-MSE | 6.5 GB | 8.7655 | 75.00% (123/164) | 68.9 — duel vs stock Q6 is 0/3 (5 tasks, HE p=1.00): the +3pp is NOISE at n=164, size win (−0.7 GB) is real, verdict win is not |
| MERNIK-6500-SMSE | 6.5 GB | 8.7275 | 72.56% (119/164) | 67.1 — LOSES to MSE-6500 by 2.4pp: PPL tie, verdict fail. Refinement: SMSE had MORE Q8 (125 vs 108) yet lost — at big budgets the MIDDLE decides (MSE Q6 67 vs SMSE 41), not penthouses |
| Q6_K (stock ref, not uploaded) | 7.2 GB | 9.1362 | 71.95% (118/164) | 66.5% — plain flat Q6_K, anyone can reproduce (bartowski-equivalent); reference only |
| Q4_K_M (stock ref, not uploaded) | 5.63 GB | 8.6671 | 70.73% (116/164) | 66.5% — measured 2026-09-29, our harness; verdict ties our 5.0 GB builds (115-117, ±2 tasks, noise) at +0.6 GB: same verdict, smaller file |
| Q3_K_M (stock ref, not uploaded) | 4.62 GB | 13.4014 | 18.29% (30/164) | 17.7% (29/164) — measured 2026-09-29, our harness; our 4.4 GB 4500-SMSE beats it 90 vs 30 (+60 tasks) at equal weight |
Files prefixed *-SMSE.gguf; the first two shipped as *-BALANCE.gguf (same formula, old name).
Reference duel (same harness): OxCoder-SMAPE-5100 88.41%, NeoHorse-SMAPE-5100 82.93%.
How to read our HE column
We use HumanEval for comparison, not for score. MiMo's home arenas are agentic (SWE, terminal, tools — see the base model card); HE is just the probe. The rule: if a small hybrid scores roughly like stock Q6 on the same harness, most of its capabilities likely survived quantization — the verdict column certifies preservation, not rank.
Why this over stock quants
Built on MiMo's agentic + vision foundation (27% visual SFT on top of Qwen3.5's native early-fusion vision) — the hybrid allocation keeps boundary attention at high precision and norms out-of-budget (F32 in file by the 1D law), so the small builds hold together better than flat stock quants:
| Build | Size | PPL (ctx1024) | HE pass@1 |
|---|---|---|---|
| MERNIK-5100-SMAPE (this) | 5.0 GB | 8.7435 | 70.12% |
| Q6_K (stock, bigger!) | 7.2 GB | 9.1362 | 71.95% |
Honest split, three columns: the 5.0 GB hybrid beats the 7.2 GB stock on perplexity (8.74 vs 9.14 — a real gap); the verdict column is a tie within noise (115 vs 118, 3 tasks — this instrument cannot resolve below ~13 tasks at n=164). Allocation wins on size (−2.2 GB for a tied verdict). The verdict race at 4500/6500 is decided (table above): the only verdict gap that clears the floor is SMSE-4500's +41-task rescue over SMAPE (3/3 columns, p<0.0001).
Stock Q4_K_M measured 2026-09-29 (see table); Q3_K_M verdict in the table. No unmeasured reference rows remain.
Usage
llama-server -m MiMo-V2.6-Distill-Qwen-9B-MERNIK-5100.gguf --port 28082 -ngl 99 -c 8192 --jinja
The model ships its own jinja template (macro-based, not the classic Qwen one) — --jinja works natively.