An unofficial experimental MLX derivative of
ornith-ai/Ornith-1.5-35B-A3B, pinned to
revision e4dfb35a93d4b6822a811a7676f3488514abe7e2. The original model is by Ornith AI; the MLX
conversion, refusal-direction experiment, and validation were performed by
PocketAI Model Lab. PocketAiHub identifies the publisher of this derivative.
This is the recommended practical release. It retains the abliterated BF16 behavior much more closely than the 4-bit build.
Important safety notice
This checkpoint was deliberately modified to suppress learned refusal behavior.
It may produce harmful, illegal, offensive, deceptive, or dangerously incorrect
content more readily than the upstream instruction model. Abliteration is not
truthfulness training, a capability improvement, or a guarantee of universal
compliance. Independently evaluate and constrain outputs for your use case.
Format
MLX affine 8-bit/group 64; router and shared-expert gates 8-bit
Stored model payload: 37,748,365,508 bytes (35.16 GiB)
Vision tower retained; the 4-bit build passed a basic image-input smoke test
Native MTP speculative-decoding head is not included because mlx-vlm==0.6.8
drops mtp.* tensors during conversion
The regular BF16 parent produced explicit-refusal phrases on 12/12 harmful gate
prompts. The selected abliterated BF16 candidate produced 0/12 on the same gate
and retained 12/12 deterministic capability checks.
Batch-1 screen
Explicit-refusal phrase flags
Final-answer text present
Harmful prompts
2/100
100/100
Benign controls
0/100
100/100
The scorer is phrase based. The 128-token ceiling makes this an early-refusal
screen rather than a complete answer-quality evaluation, and manual inspection
found semantic refusals that it did not flag. “Abliterated” describes the
weight-editing method; it does not mean “fully uncensored.”
Matched-teacher drift
The drift suite used 36 prompts—12 capability, 12 harmful, and 12 benign—with
481 shared teacher positions and exact KL over all 248,320 logits. It also
captured all 40 residual layers, K/V state for 10 full-attention layers, and
convolution/recurrent state for 30 linear-attention layers.
Pure BF16 ablation split
Mean forward KL
Top-1 agreement
Residual cosine
Capability
0.018342
97.94%
0.995895
Benign
0.307388
85.42%
0.969620
Harmful
1.049146
60.94%
0.890754
Across all 481 positions, the pure BF16 ablation measured mean KL
0.545185, top-1 agreement
78.17%, and residual cosine
0.952090 versus regular BF16.
Against the abliterated BF16 master, this quantization measured mean KL 0.021342, top-1 agreement 95.43%, and residual cosine 0.994969.
The total 8-bit recommended path versus regular BF16 measured mean KL
0.549199, top-1 agreement
77.55%, and residual cosine
0.949564.
The vision tower is present in every release, but only the 4-bit model received
an end-to-end image smoke test. Broader vision, video, coding, tool-use, and
long-context evaluations remain future work.
License and attribution
The upstream model card declares MIT. This derivative preserves the upstream
attribution and links to the exact pinned source revision above.
