license: apache-2.0
library_name: mlx
pipeline_tag: image-text-to-text
base_model:

• SC117/occamy-1.0-abliterated-FIT-GGUF
• Accio-Lab/occamy-1.0
• KaedeTai/Ornith-1.5-35B-A3B-BigBang-MTP-zh-mlx-4bit
• Qwen/Qwen3.6-35B-A3B
  language:
• zh
• en
  tags:
• mlx
• apple-silicon
• mtp
• speculative-decoding
• qwen3_5_moe
• moe
• vision-language
• abliterated
• omlx
• 4-bit

Occamy-1.0 abliterated, back from GGUF to MLX, with a draft head it never had

SC117/occamy-1.0-abliterated-FIT-GGUF
is the abliterated Accio-Lab/occamy-1.0,
and it exists only as GGUF. This repo converts its BF16 GGUF back to the
Hugging Face layout, quantizes it for MLX the same way as
KaedeTai/Ornith-1.5-35B-A3B-BigBang-MTP-mlx-4bit,
puts Occamy's vision tower back, and adds a multi-token-prediction (MTP) draft
head taken unchanged from
KaedeTai/Ornith-1.5-35B-A3B-BigBang-MTP-zh-mlx-4bit.

A head trained on a different fine-tune of the same base drafts for Occamy as
well as it drafts for its own trunk: 65.6% of drafts accepted, against 64.9% on
Ornith.

Same prompt set, oMLX Lightning MTP, 4 draft tokens, read from oMLX's
MTP[n] ... accept=a/d log lines. Speculative decoding is verify-then-accept, so
the head changes speed, not what the model knows.

What is in the files

About 20 GB. config.json carries 92 per-module overrides: 80 for the router
gates and 12 for the head. Abliteration does not touch the vision tower, so it
comes straight from Accio's release.

Getting back from GGUF

llama.cpp's QWEN35MOE conversion rewrites several tensors on the way in.
scripts/convert_occamy.py undoes each step before the usual HF→MLX pass:

• zero-centred RMSNorms were stored +1 → subtract 1 (linear_attn.norm excepted)
• A_log was stored as ssm_a = -exp(A_log) → log(-ssm_a)
• dt_bias was renamed ssm_dt.bias, and conv1d was squeezed → restored to [C, 1, K]
• linear-attention V heads were reordered from grouped to tiled order across
  in_proj_qkv, in_proj_z, in_proj_a/b, A_log, dt_bias, conv1d and
  out_proj → inverse permutation
• per-expert matrices were stacked into ffn_*_exps → kept stacked, which is
  what MLX's switch_mlp wants

Check against the unabliterated base: 42 reconstructed tensors from two
layers (one linear-attention, one full-attention) were compared with
Accio-Lab/occamy-1.0. 40 are bit-identical. The other two are exactly the
matrices abliteration edits: self_attn.o_proj (relative difference 1.7e-2)
and one expert's down_proj (5.9e-2). A wrong inverse transform could not
leave the other 40 identical. The GGUF's SHA-256 matched SC117's SHA256SUMS
before conversion.

General knowledge: TMMLU+

Evaluated on the trunk, without MTP. 3,334 questions from
ikala/tmmluplus: 67 subjects,
up to 50 each, sampled with seed 0. Zero-shot, thinking off. The score is the
mean of the four group means, and the confidence intervals are bootstrapped
within subjects.

On the same questions, occamy scores +2.1 points over Ornith (paired
bootstrap, 95% CI +0.8 to +3.4). Most of the gain is in humanities. Two answers
out of 3,334 could not be parsed and count as wrong.

Serving it in oMLX

• Let oMLX register the directory itself (open it once in the app), then set
  mtp_enabled: true, mtp_num_draft_tokens: 4 and enable_thinking: false
  through the app or the admin API. A hand-written entry in
  ~/.omlx/model_settings.json is not picked up. Without that entry,
  enable_thinking defaults to true.
• Do not set vlm_mtp_enabled. It is a different feature, mutually
  exclusive with mtp_enabled, and it silently switches this head off.
• When it works, every request logs MTP path activated … (model has mtp_forward …)
  and, at the end, MTP[0] finish=… accept=a/d (…%).
• With MTP on, outputs are not bit-identical to MTP off. oMLX is not
  bit-reproducible run to run even on the donor model, so this is batching and
  kernel order, not the head.

One tensor, one file

The head lives only in model-mtp.safetensors. Occamy ships no MTP tensors, so
no shard duplicates the head (a hazard described in the donor repo): 2,134
tensors in the index, 44 of them language_model.mtp.*, each stored once.
scripts/graft_mtp.py is the whole graft. It copies the donor's
language_model.mtp.* tensors, text_config.mtp_num_hidden_layers and the
head's quantization overrides.

Accio publishes an experimental MTP head of its own,
Accio-Lab/occamy-1.0-MTP. It
was not tried here.

Limitations

• Abliterated. Refusal directions were removed by SC117. Refusal behaviour
  was not measured for this repo; see SC117's card. You are responsible for how
  you use it.
• The draft head was trained on one deployment's Traditional Chinese traffic
  (health consultation and customer service) on a different trunk. On other
  workloads the acceptance rate may be lower.
• The numbers above are acceptance rates, not a paired tokens-per-second
  benchmark.
• Everything here comes from one machine (M5 Max, 128 GB) with oMLX and MLX.
  Nothing was tested with vLLM or transformers.

Scripts

• scripts/convert_occamy.py GGUF BASE_META_DIR DST [--check] builds the MLX
  weights from the GGUF. --check runs the comparison against the base repo
  instead. The vision tower comes from BASE_META_DIR/model-visual.safetensors,
  or from the base repo over HTTP range requests (scripts/hfrange.py).
• scripts/graft_mtp.py TARGET DONOR OUT adds the head. It hard-links the
  trunk files, so the graft costs no extra disk.

Credits and license

• Accio-Lab trained Occamy-1.0
  (Apache-2.0) from Qwen/Qwen3.6-35B-A3B
  (Apache-2.0).
• SC117 did the
  abliteration (Apache-2.0).
• The MTP head comes from the MIT-licensed Ornith lineage:
  EryriLabs/Ornith-1.5-35B-A3B-BigBang-MTP,
  retrained for Chinese in
  KaedeTai/Ornith-1.5-35B-A3B-BigBang-MTP-zh-mlx-4bit.
• oMLX provides Lightning MTP.
• The conversion reads llama.cpp's conversion/qwen.py backwards.

The repository is released under Apache-2.0 (see LICENSE, Accio's original).
The MTP head's MIT terms are compatible with that.