license: mit
library_name: mlx-vlm
pipeline_tag: image-text-to-text
base_model:

• huihui-ai/Huihui-Ornith-1.5-35B-A3B-abliterated
  base_model_relation: quantized
  tags:
• mlx
• oq
• quantized
• abliterated
• uncensored
• mtp
• qwen3_5_moe
• moe
• vision
• apple-silicon

Ornith-1.5-35B-A3B Abliterated — oQ4e + MTP (MLX)

oQ4e mixed-precision MLX quantization of
huihui-ai/Huihui-Ornith-1.5-35B-A3B-abliterated
(~4.9 bpw, ~20 GB), with a trained MTP speculative-decoding head grafted from
KaedeTai/Ornith-1.5-35B-A3B-BigBang-MTP-mlx-4bit.
Vision tower preserved.

Ornith-1.5-35B-A3B is a hybrid GatedDeltaNet + attention Mixture-of-Experts
(256 routed experts, 8 active, ~3B active params) vision-language model. This
build targets Apple Silicon via mlx-vlm / oMLX and decodes at ~67 tok/s
stock, 76–81 tok/s with Lightning MTP enabled on an M4 Pro 64 GB.

⚠️ Uncensored research artifact. The abliterated upstream checkpoint
suppresses refusal behavior in the weights; this quantized build inherits
that property. It will answer requests a stock model declines. Do not
expose it as a public endpoint without your own moderation layer. You are
responsible for lawful use.

Lineage

Validation

Self-reported, measured on this quantized checkpoint on a Mac mini M4 Pro
(64 GB), mlx_vlm unless noted. Script: run_validation.py + results in
validation-summary.json in this repo.

Refusal screen methodology

100 harmful prompts (mlabonne/harmful_behaviors test split — the same set
most abliteration cards report against) + 100 benign controls
(mlabonne/harmless_alpaca). enable_thinking=false, temp 0, 256-token cap,
Heretic-style keyword scorer.

The keyword count is an upper bound. Most flagged harmful outputs were
compliant answers that merely discussed flagged words (insider-trading
explainers, framed hacking guides, drug-manufacturing overviews — often with
disclaimers or educational framing). Genuine refusals after manual review:
~2–4/100 (soft reframes giving background instead of the requested
procedure). This is consistent with the upstream ablation's shallow layer
coverage — it de-censors less aggressively than Heretic-style builds, and in
exchange likely perturbs capability less.

MTP: +14–20% decode speedup (head matters)

The grafted BigBang 4-bit head gives a measured +14% (essay) / +20% (code)
decode speedup on an M4 Pro (same-session A/B, Lightning MTP on vs off),
with ~66% draft acceptance (62.6–69.9% across runs) logged by the
inference server.

Notably, the first head we grafted — the bf16 shisa-ai MTP-ONLY head
distilled against the stock trunk — measured ~8% slower than off on
this ablated checkpoint (70.5 vs 76.6 tok/s): draft acceptance drops when the
donor's training distribution doesn't match the quantized/ablated weights.
The BigBang head was trained on a tuned Ornith derivative and accepts
noticeably better here. If you reproduce this build, donor choice is the
difference between a positive and negative speedup.

Usage

Standard MLX safetensors layout — loads in mlx-vlm (keeps the vision tower;
mlx-lm would strip it).

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template

model, processor = load("terry2010/Ornith-1.5-35B-A3B-abliterated-MLX-oQ4e-mtp")

prompt = apply_chat_template(
    processor, model.config, "Your prompt",
    num_images=0,
    enable_thinking=True,   # False skips the reasoning preamble
)
print(generate(model, processor, prompt, max_tokens=512))

Via any OpenAI-compatible server (oMLX, LM Studio):

{"chat_template_kwargs": {"enable_thinking": false}}

Thinking note: the model emits a <think> trace by default. In a quick
spot check, disabling thinking produced a wrong answer on an edge-case
numeric comparison that the thinking trace resolved correctly — leave it on
unless latency matters more than accuracy.

Vision: pass --image (mlx_vlm) or an image input through your server;
the tower is inside the checkpoint, no separate projector file needed.

Files

• model-0000{1..5}-of-00005.safetensors — oQ4e quantized weights (~20 GB total)
• model-mtp.safetensors — grafted BigBang MTP draft head (4-bit, 44 tensors)
• run_validation.py — self-contained validation script (mlx_vlm, zero pip deps)
• oq_imatrix_report.json — quantizer's own calibration/allocation report
• validation-summary.json — raw validation output incl. flagged refusal outputs
• tokenizer / preprocessor / chat template files from the upstream checkpoint

Not measured

• Benchmark scores (MMLU, HumanEval, SWE-bench) — capability-retention claims
  beyond the checks above are not made here
• Refusal rate with thinking enabled
• Multi-turn refusal re-assertion, system-prompt steering, vision-path refusals
• KL divergence vs. the abliterated parent (see oq_imatrix_report.json for
  the quantizer's own fidelity record)

License & attribution

MIT, inherited from the upstream chain. Model lineage: ornith-ai (base) →
huihui-ai (abliteration) → this build (oQ4e quant + KaedeTai BigBang
MTP head graft). All upstream licenses and acceptable-use terms apply.