Qwen3.5-4B MTPLX Optimized Speed

4-bit (group 64) Qwen3.5-4B with a calibrated native-MTP draft head, built for MTPLX on Apple Silicon. 2.47 GB on disk, ~2.9 GiB peak at load. The fastest model MTPLX ships.

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

• AR baseline: 133.6 tok/s
• MTP depth 3: 227.8 tok/s (1.71x), first-position acceptance 0.95

The engine reads the tuned depth from mtplx_runtime.json; no manual configuration needed. Runs on any Apple Silicon Mac with 8 GB+ of unified memory.

July 2026 rebuild (important)

Earlier revisions of this repo shipped a defective MTP sidecar: the RMSNorm weights were stored in the raw zero-centered convention and never restored, so the draft head proposed garbage and MTP made the model slower than plain decoding (reported as #176, thanks lBroth). This revision is rebuilt from source with the fixed forge. MTPLX 2.2.0+ also detects and heals the old sidecar at load, so existing downloads recover without re-downloading.

Usage

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

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