license: mit
base_model: huihui-ai/Huihui-Ornith-1.5-35B-A3B-abliterated
tags:
- mlx
- mtplx
- qwen3-moe
- ornith
- abliterated
- uncensored
- fp16
- apple-silicon
pipeline_tag: text-generation
Huihui Ornith 1.5 35B A3B Abliterated — MTPLX mixed 8/4-bit FP16
MTPLX (native MTP speculative decoding on Apple Silicon) forge of
huihui-ai/Huihui-Ornith-1.5-35B-A3B-abliterated,
itself an abliterated (refusal-removed, layers 11–29) uncensored build of
ornith-ai/Ornith-1.5-35B-A3B.
Quant recipe (stolen from the attention8 Ornith forge, fp16 lane)
- body: 4-bit affine, group 32 — MoE experts (bulk, sparse-active)
- 8-bit affine, group 64: all
linear_attnprojections, allself_attnq/k/v/o,mlp.gate,mlp.shared_expert_gate,lm_head,embed_tokens - trunk dtype fp16 + MTP sidecar fp16 — the fast native path on M1/M2 Macs,
which have no native BF16. Runs fine on M3+ too.
Measured on M1 Max 64GB
| mode | tok/s | vs AR |
|---|---|---|
| AR | 51.1 | 1.00x |
| D1 | 68.2 | 1.34x |
| D2 | 74.8 | 1.46x |
| D3 | 73.3 | 1.44x |
Tuned best: D2 (74.0 t/s, 1.49x). Profile: sustained.
Use
mtplx serve --model r3lax/Huihui-Ornith-1.5-35B-A3B-abliterated-MTPLX-mixed8-4bit-FP16 --download
License: MIT (matches source). Forged with MTPLX 2.11.2.