davidau-catalog-verdicts

DavidAU (huggingface.co/DavidAU) — what's worth pulling, and what isn't

Scanned the full profile 2026-07-26 looking for something we don't already

have. Conclusion: nothing in the catalog beat what we run. Most of his ~40

recent repos are creative/horror/roleplay merges (off our coding/agentic/vision

lane — writer/qwythos covers creative) or things we already serve. Three

plausible candidates were evaluated in depth; all three REJECTED. Don't

re-evaluate these without a specific new reason.

The three verdicts

1. OpenAi-GPT-oss-20b-abliterated-uncensored-NEO-Imatrix (Q5_1 CODEPlus, 15.7GB)

Served as coder-uncensored, a deliberate sampler-clone of [[gpt-oss-20b]]

(coder) so weights were the only variable. REJECTED — grill killed

mid-run once it was obviously degraded: failed articulation_points (which

plain coder PASSES) plus multiple runaway 16k-token-cap generations

(expr_eval, articulation_points). Abliteration (huihui refusal-ablation)

degraded reasoning into looping. Also not a space saving: gpt-oss ships

MXFP4-native MoE weights, so "quantizing" it quantizes huihui's BF16 *upcast* —

Q5_1 lands 15.7GB, ABOVE coder's 13.8GB F16. You pay 2GB for abliteration.

2. GLM-4.7-Flash-Grande-Heretic-UNCENSORED-42B-A3B (Q4_K_M, 24GB)

Brainstorm-20x layer expansion (67 layers/1186 tensors vs base 48/844) of

GLM-4.7-Flash-heretic. Same deepseek2/MLA arch as [[glm-4.7-flash]], A3B so

still MoE-class. Served as glm-grande, sampler cloned from glm-flash.

REJECTED — grill 17/23 vs glm-flash's 20/23

(bench/results-glm-grande-20260726-144849.log): r1 5/8, r2 2/5, r3 5/5,

r4 5/5. All 3 lost points were on the algorithmic rounds and **all 5 fails were

16k-token-cap RUNAWAYS** (expr_eval, wildcard_match,

weighted_interval_scheduling, bounded_blocking_queue, text_justify) —

exactly the card's own warning ("May loop / spit out odd characters"). On

weighted_interval_scheduling, which glm-flash also fails, glm-flash gives a

*clean wrong answer* while Grande loops to cap: **the expansion converts bounded

misses into runaways.** Dominated on every axis — worse coding, slower (~50-62

vs 80-90 t/s), larger (24 vs 17.5GB), verbose, capped at 131k ctx (VRAM-maxed,

~500MB free/card) vs glm-flash's 198k — for the *same* tool/agentic ability

(r3+r4 were a matching 10/10). Rounds 3-4 are where the shared GLM lineage still

shines; the expansion only hurt the algorithmic reasoning.

3. Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-NEO-CODE

Not grilled — rejected on feasibility. qwen35 arch (same as [[fable-711-27b]]/

tess, so it WOULD load), 256k trained, vision (mmproj). But DENSE ~39B (card:

"dense, not moe", 96 layers, Brainstorm-expanded from Qwen3.6-27B) and the repo

ships NO mtp- draft head. Dense 27B tess runs 14.3 t/s without MTP; a dense

39B scales to ~10 t/s — [[kimi-dev-72b]] rejection territory — and it's a

*thinking* model burning long traces at that speed. Fits VRAM at IQ4_XS (20.6GB)

/128k, but redundant with fable-fusion's niche (uncensored+vision+coding, 23/23,

31 t/s w/ MTP, 256k). Would only pull for max-quality uncensored reasoning if

~10 t/s were acceptable.

The generalizable pattern (this is the reusable lesson)

• DavidAU Brainstorm layer-expansions add instability, not intelligence on

this box's grill. Both expansions (Grande-42B, Deckard-40B) lose to their

smaller base: expansion inflates size + slows decode + (Grande) causes

loop-to-cap runaways. The base model is the better pick every time so far.

• Abliteration/heretic variants degrade reasoning into looping at the base's

proven sampler. If ever re-tried, the card's higher rep-pen (1.02-1.1) is the

first lever, but the ceiling is still below the un-abliterated base.

• gpt-oss quants never save space (MXFP4-native → any GGUF quant upcasts).
• A controlled A/B is the right method: clone the base entry's sampler/KV so

weights are the only variable, hand-start on a spare port, grill the port

directly (llama-swap won't swap). Grade PER-TASK vs the base, and treat a

16k-cap hit as a *runaway* (looping), a categorically worse failure than a

clean wrong answer. Same method that rejected [[genesis-hermes-v5]].

• What we DO already have from DavidAU and kept: [[fable-711-27b]]

(fable-fusion), [[glm-4.7-flash]] (glm-flash). Those are the wins; the pattern

above is why the newer/bigger variants aren't.