MLX 4-bit build of DavidAU's Qwen3.6-27B Fable-Fusion-711 tune, packaged as a native MTPLX
artifact: working multi-token-prediction head (bf16 sidecar + calibrated
runtime contract) for exact speculative decoding, plus the vision tower for
image input. 16 GB.
This is the first MLX conversion of Fable-Fusion in any form.
Behavioral correction — July 26, 2026
This conversion faithfully reproduces the released Fable-Fusion weights, but
those weights do not reproduce the source card's 4/100 refusal claim. On
the exact Heretic KeywordRate protocol (same 100 prompts, system prompt,
100-token cap, and 33 substring markers), official Qwen scored 99/100,
DavidAU's released Q6_K scored 81/100, and this converted tune scored
79/100. All outputs were non-empty.
KeywordRate is a refusal-marker proxy, not a semantic compliance judge.
Accordingly, this repository is no longer described or tagged as uncensored.
For the later official-Qwen reduced-refusal artifact, see
Qwen3.6-27B-Refusal-Ablation-v2-MTPLX-6bit.
Decode throughput receipt — M5 Max (128 GB)
Measurement revision, 2026-07-30. The earlier single-sample figures were
plausible observations, but they under-characterized an output-dependent
distribution. A much lower correction posted on 2026-07-28 was itself wrong
because that run was load-confounded; it has been withdrawn. The
condition-instrumented receipt below supersedes both for this exact protocol.
These are decode rates for one pinned long-code prompt and sampler on one
Apple M5 Max. They are not universal rates for the model, every prompt, or
every serving stack.
Default MTP is bimodal on this workload
The default MTPLX runtime is a hybrid: it can use ordinary MTP drafting and a
separate prompt-context-copy path. The 30-trial sample at each depth split into
two clearly separated throughput clusters:
Mode
Low cluster: mean; range (trials)
High cluster: mean; range (trials)
Observed high-cluster frequency
MTP-only control, context copy off: mean; range (trials)
D1
51.17; 48.06–62.59 tok/s (18)
79.23; 68.71–82.49 tok/s (12)
12/30 = 40.0%
49.36; 47.82–51.04 tok/s (10)
D2
58.88; 56.40–62.00 tok/s (18)
90.69; 79.88–95.92 tok/s (12)
12/30 = 40.0%
58.71; 54.82–61.93 tok/s (10)
D3
66.11; 60.76–77.10 tok/s (19)
94.12; 81.92–100.79 tok/s (11)
11/30 = 36.7%
64.09; 60.85–67.10 tok/s (10)
The six same-seed AR sentinels bracketing the expanded run measured
30.35–30.76 tok/s. The before/after token streams were identical for each
seed.
Conditions were captured around every value above:
Mode
1-minute load-average envelope
Thermal states observed
AR sentinels
1.92–2.63
nominal, fair
D1, both arms
1.92–3.53
nominal, fair
D2, both arms
1.73–3.59
nominal, fair
D3, both arms
1.62–3.59
nominal, fair
There is deliberately no single “default D1/D2/D3” number here. On these
samples, the high cluster coincided with the prompt-context-copy path accepting
large blocks: 112–137 copied tokens at D1, 108–137 at D2, and 127–137 at
D3. MTPLX matched generated context to prompt history, proposed a prompt
block, verified it in one target pass, committed the accepted tokens, and
bypassed the corresponding ordinary MTP draft/verify cycles. Disabling context
copy removed every high observation.
Small copy events also occurred in the low cluster: context copy accepted at
least one token in 25/30 trials at every depth. Therefore 37–40% is the
observed frequency of the copy-heavy high-throughput cluster, not the
frequency with which context copy did anything. Both frequencies are specific
to this prompt, sampler, and seed set.
Draft acceptance
These are ordinary-MTP positional accepted/drafted ratios. They exclude copied
prompt tokens, which have separate counters. Each entry is
pooled ratio [minimum–maximum per trial].
The high-rate regime is therefore not a second ordinary-MTP acceptance mode;
accepted prompt-copy blocks displace ordinary MTP work.
Exact measurement scope
The measurement tool was the local
probe_fable711_4bit_mtp_distribution.py harness, which calls
mtplx.generation.generate_ar and generate_mtpk directly rather than
mtplx tune. The recorded command was:
Held fixed: the checkpoint and hashed MTP runtime contract; the single
long_code_uncapped prompt; chat template with thinking disabled; temperature
0.6, top-p 0.95, top-k 20; 256-token cap; 24-token warmup per arm/depth;
performance-cold; persistent MTP cache; committed MTP history;
capture_commit; and linear-gdn-from-conv-tape. Default-runtime seeds were
0–29 at D1/D2/D3; copy-disabled controls were seeds 0–9, interleaved. One
default D1 trial stopped normally at 241 tokens in the low cluster; the other
119 MTP trials reached 256 tokens.
The completed run used macOS 27.0, MLX 0.32.0, MTPLX 2.3.0, and Python 3.13.14
on AC power. Conditions were captured immediately before and after every
trial: 1-minute load average was 1.62–3.59; no known competing GPU
benchmark was detected; and NSProcessInfo.thermalState ranged from
nominal to fair. The stable same-seed AR sentinels did not show a coherent
system-wide slowdown. Fan RPM was not determinable, and actual per-process
GPU attribution was not determinable without privileged telemetry.
Primary evidence receipts:
fable711_4bit_mtp_distribution_protocol_v2.json,
fable711_4bit_mtp_distribution_raw_v2.json, and
fable711_4bit_mtp_distribution_analysis_v2.json. Their SHA-256 values are,
respectively, 36eaae0a…, a5bd7db5…, and d170be6c…; the structural
validator passed 27/27 checks.
Other measured checks
Check
Result
AIME quick bench (5 problems)
5/5
Executable coding eval (8 tasks, tests run)
8/8
Perplexity — wikitext-2, ctx 2048
5.9499
Perplexity context, same harness throughout: the bf16 reconstruction measures
5.7525, so this quantization costs +0.1974, a real degradation well outside
the ±0.034 measurement uncertainty. That is the price of 4-bit; the 6-bit
build is indistinguishable from bf16 if you can spare the memory. DavidAU's
Q8_0 GGUF measures 5.6635 under llama.cpp's own methodology (different
tokenizer/BOS handling, so treat it as a sanity anchor rather than a directly
comparable number).
Fable-Fusion-711 / NM are weight-level: this is DavidAU's exact tune,
dequantized from his own Q8_0 export. The source repository includes
Uncensored-Heretic in its name, but that behavioral claim did not survive
the released multi-stage tune/merge.
MTP is his tune's own head, kept at bf16 — higher precision than the
Q8_0 MTP tensors in the GGUF — then forge-calibrated and verified at depths
1–3.
NEO / MAX do not apply. NEO-Imatrix shapes his sub-8-bit GGUF quants
only; the Q8_0 source doesn't use it, so nothing was lost taking this path.
MAX (16-bit output tensor) is GGUF packaging — here lm_head sits at body
precision.
Vision works (tower grafted from mmproj-BF16.gguf, verified on real
images).
How this was made (and the traps if you try it)
DavidAU published GGUF only, so this was reconstructed from his
highest-precision export — the Q8_0 MTP GGUF — by inverting every transform
llama.cpp applies when it converts Qwen3.5/3.6. Get any of these wrong and the
model degenerates while still passing consistency-style checks:
Zero-centered RMSNorm baking. GGUF stores these norms as weight + 1:
input/post-attention layernorms, q/k norms, the final norm, and all mtp.*
norms (nextn.enorm / hnorm / shared_head_norm). mlx-lm's qwen3_5
sanitize adds +1 itself, so a naive conversion double-shifts. Gated
linear_attn.norm and the vision norms are NOT baked.
Gated-DeltaNet value-head reorder (_LinearAttentionVReorderBase in
llama.cpp's conversion/qwen.py). V heads move from HF grouped order to
ggml tiled order in in_proj_qkv (V rows), in_proj_z, in_proj_a,
in_proj_b, A_log, dt_bias, conv1d (V channels), and out_proj
(columns, not rows). Note the first V head occupies the same offset in
both orderings — spot-checks at offset 0 are blind to this bug.
ssm_a = -exp(A_log) — recover with log(-x), then apply the reorder.
MTP head mapping: blk.<last>.nextn.{eh_proj,enorm,hnorm,shared_head_norm} →
mtp.{fc,pre_fc_norm_embedding,pre_fc_norm_hidden,norm}, and the final block's
attention/MLP → mtp.layers.0.*.
Every tensor was shape-checked against Qwen/Qwen3.6-27B and the layout
verified by cosine similarity against base values (a finetune leaves many
tensors at or near base, so a residual permutation shows up as ~0 correlation).
All families landed at 0.99–1.0000 after inversion.
Independently validated. After these builds were published, an F32 copy of
the same tune surfaced:
nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-F32
(112 GB, mergekit nuslerp). It has the identical structure — 1199 tensors,
15 mtp.*, 333 vision — and this reconstruction matches it at cosine
0.99997 on mtp.fc.weight and 1.00000 on
model.language_model.norm.weight. The residual is precisely the Q8_0 step
this copy came through, so the inverse transforms above are confirmed correct
against a source never used to derive them.
If you want the highest-fidelity starting point for your own quants, prefer
that F32 repo over this reconstruction — it never passed through a quantizer.
Credits and license
DavidAU — the Fable-Fusion-711 tune and
the Q8_0 MTP export this is reconstructed from. Please read
his model card
for the tune's capabilities, sampler guidance, and class notes.
