genesis-hermes-v5

Genesis-Hermes-V5 (APEX) — tested and REJECTED 2026-07-26

LuffyTheFox/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V5-GGUF, APEX quant

23.9GB. Served briefly as genesis-v5 on port 9115; entry and weights removed

the same day (commit be044ca5). Grill log kept:

bench/results-genesis-v5-20260726-124041.log.

Why it was worth a controlled test: its base_model is

HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive — *exactly* what the

uncensored entry already serves. V5 = that base + two author edits, both

shipped with zero benchmarks:

1. a graft of ~2k blocks from two FFN expert tensors taken from a

hermes-qwen3.5-35b-a3b finetune (a cross-generation graft onto a 3.6

base — not a safe operation), claimed to improve function calling;

2. "Genesis", an author-invented SVD "denoise" applied to the quantized GGUF.

The A/B was clean by construction: V5's embedded chat template is

byte-identical to the base's apart from the system-guard line, so it reused

chat-templates/qwen3.6-35b-a3b-uncensored-aggressive-iq4nl-128k.jinja

verbatim. Same 131072 ctx, q8_0 KV, 50/50 split, same sampler, --no-mmproj to

match the baseline's text-only serving. **The one uncontrolled variable — quant

tier — favoured V5** (APEX 23.9GB vs IQ4_NL 19.8GB), which is what makes the

negative result conclusive.

Result: 21/23 (r1 7/8, r2 4/5, r3 5/5, r4 5/5), 101.6k tokens, 22.1 min.

• On the 11 tasks the baseline also ran: baseline 10/11, V5 9/11, and V5

spent ~2× the tokens (20,384 vs 10,483 on the three shared round-2 tasks).

• It lost weighted_interval_scheduling (wrong DP answer, not a format error)

which the *smaller* quant passes, while carrying a 4GB size advantage.

• Rounds 3+4 perfect — but that is where most competent models here score full

marks, so it is not evidence the Hermes graft did anything.

• Verdict: dominated. [[kimi-distill]] takes the same 21/23 at 18.9GB and

fewer tokens. No sign the SVD "denoise" or the cross-generation graft bought

anything; the size advantage bought verbosity, not correctness.

Method note worth reusing: the paired baseline re-grill is not optional.

The plan started out reusing uncensored's 2026-07-09 log. That log turned out

to be (a) truncated — 8 round-1 tasks, only 3 of 5 round-2, no rounds 3/4 — and

(b) not reproducible on today's build: the re-run FAILS articulation_points,

which it passed in July. Cross-build grill comparison is unsound; always re-run

the baseline on the current binary. Partial re-run kept as

bench/results-uncensored-20260726-130340.log.

See [[llama-swap-setup]] for the sudo-free kill -HUP reload used throughout.