AX ENGINE / Performance
Performance.
Version-pinned benchmark snapshots with source artifacts. The September 21 campaign compares six Qwen and Gemma workloads; historical qualification and campaign tables follow. Results depend on the pack, decoding path, prompt suite, and host.
Measured on one Mac. Shown in context.
M5 Max 128 GB · 2026-09-21 · Campaign evidence, not certification
Decode tok/s · 256 tokens · greedy · median of 20 runs (4 prompts × 5 reps)
Prefix cache off · n-gram stacking off · AX Engine 7.5.3 @ c5d22c4a
AX / MTPLX: MTP depth 3 · OMLX: depth 1 · mlx-lm: direct AR
AX / MTPLX: MTP depth 3 · mlx-lm: direct AR · OMLX could not load these raw packs.
Community packs, not AXQ. All measured lanes are direct AR; no Assistant-MTP sidecar.
One campaign host, not the M4 Pro qualification SKU. Correlated samples and sequential lanes allow thermal drift; gaps below 1.10× are treated as parity. Checkpoint Tier 1; MTP Tier 2 pending. These ratios are not product multipliers.
MTPLX uses --allow-unverified-model. Qwen 6-bit OMLX uses an imported snapshot. Unavailable means the tested pack and version could not run this lane, not zero throughput or universal model incompatibility.
Same-binary control: AX direct AR 30.14 vs mlx-lm 27.92 tok/s. Most of the Qwen acceleration comes from MTP.
Flash Next is experimental load/serve evidence only. The 20-run benchmark was not completed; MTP gates were not assessed. OMLX initialization was not tested. No speed comparison is published.
Mac mini M4 Pro 64 GB · 273 GB/s memory bandwidth
Qualification SKU · Mac mini M4 Pro 64 GB · 2026-09-17
| Runtime | Decode (tok/s) | Prefill (tok/s) |
|---|---|---|
| AX Engine 7.4.0 · ad999f3f | 31.05 | 120.3 |
| MTPLX 2.11.3 | 28.16 | 114.0 |
| oMLX 0.6.4 | 15.02 | — |
| mlx-lm 0.31.3 | 12.78 | — |
Apple M5 Max 128 GB · 614 GB/s memory bandwidth
Campaign host · Apple M5 Max 128 GB · 2026-09-15 (pre-correction)
| Runtime | Decode (tok/s) | Prefill (tok/s) |
|---|---|---|
| AX Engine 7.4.0 | 76.90 | 795.3 |
| MTPLX 2.11.2 | 70.62 | 686.6 |
| oMLX 0.6.4 | 38.47 | — |
| mlx-lm 0.31.3 | 27.90 | — |
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 ↗