8-bit (group 64) Qwen3.5-4B with a calibrated native-MTP draft head, built for MTPLX on Apple Silicon. 4.58 GB on disk, ~4.8 GiB peak at load. The highest-fidelity 4B MTPLX ships, with the largest MTP multiplier in the fleet.

Measured on an M5 Max (max fans, MTPLX 2.2.0, deterministic suite):

AR baseline: 87.4 tok/s
MTP depth 3: 191.7 tok/s (2.19x), acceptance 0.91 / 0.76 / 0.62
First-position acceptance: 0.95

The 8-bit trunk keeps output quality close to the BF16 reference while the calibrated draft head converts that fidelity into a 2.2x decode multiplier. The engine reads the tuned depth from mtplx_runtime.json. Runs on any Apple Silicon Mac with 8 GB+ of unified memory.

Provenance

New artifact (July 2026), forged with the fixed MTPLX forge after the 4B zero-acceptance defect (#176) was root-caused: draft-head RMSNorms in the original export are stored zero-centered and must be restored at extraction. The draft head is quantized int4 (group 64) with fc and norms kept in BF16, calibrated so acceptance matches the BF16 head.

Usage

Pick "Qwen 3.5 4B Optimized Quality" in the MTPLX app, or:

mtplx serve --model Youssofal/Qwen3.5-4B-MTPLX-Optimized-Quality