AX Engine 7.5.x ships dense 27B past the DRAM ceiling.See the numbers

AX ENGINE / Performance

Performance.

AX Engine leads the loadable MTP peers on the Qwen 3.8 27B AXQ 6-bit MTP pack: MTPLX (draft depth 3), OMLX (Lightning depth 1), and the direct-AR `mlx-lm` baseline. Numbers are 20-run medians on the same checkpoint.

Mac mini M4 Pro 64 GB · 273 GB/s memory bandwidth

Qualification SKU · Mac mini M4 Pro 64 GB · 2026-09-17

Qualification SKU · Mac mini M4 Pro 64 GB · 2026-09-17
RuntimeDecode (tok/s)Prefill (tok/s)
AX Engine 7.4.0 · ad999f3f31.05120.3
MTPLX 2.11.328.16114.0
oMLX 0.6.415.02
mlx-lm 0.31.312.78

Apple M5 Max 128 GB · 614 GB/s memory bandwidth

Campaign host · Apple M5 Max 128 GB · 2026-09-15 (pre-correction)

Campaign host · Apple M5 Max 128 GB · 2026-09-15 (pre-correction)
RuntimeDecode (tok/s)Prefill (tok/s)
AX Engine 7.4.076.90795.3
MTPLX 2.11.270.62686.6
oMLX 0.6.438.47
mlx-lm 0.31.327.90

MTP Tier 2 is still pending. Decode moves past the DRAM ceiling by emitting several tokens per weight pass; the bandwidth ceiling on the qualification SKU is reached at ~266 GB/s of weight-stream bandwidth.

Full peer comparison, methodology, and Tiel / Cyber-Tiel MXFP4 MTP rows ↗Qwen 3.8 27B benchmarks ↗