DavidAU Qwen3.8 27B TURBO Fable Cold Fusion 735-882 — MTPLX 4-bit FP16
MTPLX (native MTP speculative decoding on Apple Silicon) forge of DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU.
FP16 build for M1 / M2 Macs
M1/M2 GPUs have no native BF16 (emulated, slower prompt processing).
This build keeps all non-quantized weights and the MTP sidecar in FP16,
so prefill runs on the fast native path. On M3 and newer it runs fine too —
FP16 is simply the legacy-silicon lane. If you are on M3+, the BF16 twin
(blackfan23/DavidAU-Qwen3.8-27B-TURBO-Fable-Cold-Fusion-MTPLX-4bit) is the
matching alternative.
Recipe
- body: 4-bit affine, group 64, dtype fp16
- MTP policy: keep (sidecar cast to fp16)
- profile: sustained, tuned best depth D1 on M1 Max 64GB
Measured on M1 Max 64GB
| mode | tok/s | vs AR |
|---|---|---|
| AR | 18.4 | 1.00x |
| D1 | 24.6 | 1.34x |
| D2 | 22.4 | 1.22x |
| D3 | 23.0 | 1.26x |
Use
mtplx serve --model r3lax/DavidAU-Qwen3.8-27B-TURBO-735-882-MTPLX-4bit-FP16 --download
License: apache-2.0 (matches source). Forged with MTPLX 2.11.2.