Where each tool fits.
Both runtimes are Apple Silicon MTP engines for Qwen-class dense models. AX Engine keeps depth 3 sustained on the Qwen 3.8 27B AXQ pack and serves an OpenAI-compatible endpoint on 127.0.0.1:31418. MTPLX sustains depth 3 on the same checkpoint and exposes an OpenAI-compatible HTTP surface. The choice depends on whether you want the AX Engine batched-decode ceiling, the AX Code integration story, or a stand-alone MTP runtime.
| Consideration | AX Engine 7.4.0 | MTPLX 2.11.2 (M5) / 2.11.3 (M4) |
|---|---|---|
| Measured versions / dates | AX Engine 7.4.0: M5 Max, 2026-09-15 (pre-correction); M4 Pro, 2026-09-17 (build ad999f3f). | MTPLX 2.11.2: M5 Max, 2026-09-15; MTPLX 2.11.3: M4 Pro, 2026-09-17. |
| MTP depth stability | Sustains draft depth 3 on Qwen 3.8 27B AXQ across the 20-run median. | Sustains draft depth 3; depth-4 publish option under MTPLX 2.11.x is gated by acceptance thresholds. |
| Decode (tok/s) | 76.90 on M5 Max 128 GB · 31.05 on Mac mini M4 Pro 64 GB. | 70.62 on M5 Max 128 GB · 28.16 on Mac mini M4 Pro 64 GB. |
| Prefill (tok/s) | 795.3 on M5 Max 128 GB · 120.3 on Mac mini M4 Pro 64 GB. | 686.6 on M5 Max 128 GB · 114.0 on Mac mini M4 Pro 64 GB. |
| Endpoint shape | OpenAI-compatible HTTP on 127.0.0.1:31418; gRPC off by default; LAN advertise opt-in. | OpenAI-compatible HTTP; default port and serve path documented in the MTPLX README. |
| Multi-model serving | Multi-model co-resident on hosts that fit ≥2 packs; per-model `--port` selection. | Single-model default; multi-model supported via additional processes. |
| Tie-in to AX Code | AX Engine is the local-model provider recommended by AX Code's `/connect` flow. | Stand-alone; no documented AX Code integration. |
Reproduce the comparison.
Run the AX Engine peer benchmark against MTPLX on the same Apple Silicon host and the same Qwen 3.8 27B AXQ checkpoint. Treat the published numbers as one valid data point, not a verdict.
Apple Silicon hosts vary in sustained bandwidth, thermal envelope, and MLX kernel version. Numbers from one SKU do not transfer cleanly to another. MTPLX and AX Engine both accept the same `--host --port` and OpenAI-compatible wire format.
AX Engine GitHub repository ↗Is AX Engine faster than MTPLX, full stop?
On the published Qwen 3.8 27B AXQ 6-bit MTP comparison, yes — by ~9% on decode and ~16% on prefill on M5 Max 128 GB. The gap is smaller on M4 Pro 64 GB. Speed is host-, thermal-, and MLX-kernel-dependent; the only meaningful answer for your workload is to run both on your hardware with the same checkpoint.
What to measure.
Install both runtimes on the same Apple Silicon host, serve the same Qwen 3.8 27B AXQ 6-bit MTP checkpoint, and drive each with the same OpenAI-shaped traffic (chat completion, code completion, RAG with long context). Record throughput, time-to-first-token, and retry behavior. The `ax-engine bench` and MTPLX CLI both emit JSON artifacts you can diff directly.
Install ↗Sources
Every number on this page traces to a checked-in benchmark artifact. Re-run the comparison against your own host and load before drawing conclusions.
The honest summary.
AX Engine leads MTPLX on the Qwen 3.8 27B AXQ 6-bit MTP checkpoint on both the qualification SKU and the campaign host — by 9-10% on decode and 16% on prefill on M5 Max 128 GB. The gap is smaller on M4 Pro 64 GB. Treat one row as a snapshot, not a verdict; the peer benchmark re-runs every AX Engine minor release.
No throughput number from one host is portable to another. Verify against your hardware, your prompt mix, and the model revision you actually deploy.
Performance ↗