Hardware

Apple M4 Pro

Veizik on Apple M4 Pro — Apple Silicon · Metal · M4.

Overview

What this device is.

Apple M4 ProAvailable

Vendor: Apple Silicon
Backend: Metal
Platform: macOS on Apple Silicon (arm64)
Generation: M4
Architecture: M4
Availability: Available

On Apple M4 Pro, 2 model families are verified end to end on a published build. Evidence and availability are separate: “verified” is a first-hand run on this exact device, not a download, and “available” never means verified on your hardware.

All devices · Platform requirements · Benchmarks & method

Engines

The engine that runs here.

Veizik native engineAvailable

Backend: Metal
Architecture: M4
Model path: dep0 — pure native, no Python runtime
Families carried: Qwen2.5, Qwen3

One engine runs every model family on this backend. Each GPU architecture is carried as its own native kernel set; a family is “verified” on a device only when it has been run there end to end.

Models

What runs on this device.

Pick a device, get a straight answer — which model families are verified, carried by architecture, or in preparation.

VerifiedVerifiedRun end to end on this exact device, on a published build.Architecture-supportedArchitecture-supportedVerified on another device of the same architecture; this device itself is untested.Expected, untestedExpected, untestedThe engine targets this architecture; nobody has run this model on it. No claim is made.In preparationIn preparationNo build is downloadable yet, and no figure is offered.Not offeredNot offeredThis family is not offered on this platform.UnknownUnknownNo information yet — the platform is announced with the release.

Model family	Evidence on this device	Availability
Qwen2.5 Language	Verified Verified	Available
Qwen3 Language	Verified Verified	Available

Performance

Measured on this exact device.

Model	Task	Weights in memory
Qwen2.5-0.5B	Language	343.6 MB
Qwen2.5-1.5B	Language	926.9 MB
Qwen2.5-7B	Language	4,114.0 MB
Qwen3-1.7B	Language	968.0 MB

Deterministic load — the weights occupy the same memory every run.

Decode, tokens per second Measured here

Generated tokens per second, greedy, one request, through veizik run. Median of the runs shown (hover for the range).

Model	Mode	Workload	tok/s
Qwen2.5-0.5B	—	512 / 128 tok	462.8 tok/s · 11 runs · range 425.7–475.6
Qwen2.5-0.5B	—	3968 / 128 tok	358.4 tok/s · 11 runs · range 344.2–360.4
Qwen2.5-0.5B	—	8192 / 128 tok	353.2 tok/s · 10 runs · range 349.0–362.4
Qwen2.5-0.5B	—	16384 / 128 tok	279.0 tok/s · 11 runs · range 263.8–284.9
Qwen2.5-0.5B	—	24576 / 128 tok	228.6 tok/s · 9 runs · range 209.6–232.4
Qwen2.5-0.5B	—	30000 / 128 tok	199.8 tok/s · 10 runs · range 187.4–204.4
Qwen2.5-1.5B	—	512 / 128 tok	215.0 tok/s · 11 runs · range 199.9–216.6
Qwen2.5-1.5B	—	3968 / 128 tok	172.3 tok/s · 9 runs · range 165.0–176.9
Qwen2.5-1.5B	—	8192 / 128 tok	159.8 tok/s · 11 runs · range 156.3–163.5
Qwen2.5-1.5B	—	16384 / 128 tok	134.1 tok/s · 11 runs · range 128.3–139.5
Qwen2.5-1.5B	—	24576 / 128 tok	117.8 tok/s · 11 runs · range 114.4–119.2
Qwen2.5-1.5B	—	30000 / 128 tok	108.8 tok/s · 7 runs · range 106.7–109.3
Qwen2.5-7B	—	512 / 128 tok	54.9 tok/s · 11 runs · range 50.2–56.3
Qwen2.5-7B	—	3968 / 128 tok	47.9 tok/s · 10 runs · range 44.5–48.4
Qwen2.5-7B	—	8192 / 128 tok	47.4 tok/s · 6 runs · range 45.9–47.8
Qwen2.5-7B	—	16384 / 128 tok	42.3 tok/s · 11 runs · range 41.6–42.9
Qwen2.5-7B	—	24576 / 128 tok	38.4 tok/s · 10 runs · range 37.8–38.8
Qwen3-1.7B	—	512 / 128 tok	204.3 tok/s · 11 runs · range 200.1–207.7
Qwen3-1.7B	—	3968 / 128 tok	150.1 tok/s · 10 runs · range 144.5–152.0
Qwen3-1.7B	—	8192 / 128 tok	111.3 tok/s · 11 runs · range 108.6–112.8
Qwen3-1.7B	—	16384 / 128 tok	78.1 tok/s · 10 runs · range 75.2–78.5
Qwen3-1.7B	—	24576 / 128 tok	58.1 tok/s · 11 runs · range 56.9–58.6
Qwen3-1.7B	—	30000 / 128 tok	50.0 tok/s · 10 runs · range 47.9–50.8

Prefill, tokens per second Measured here

Prompt tokens processed per second.

Model	Mode	Workload	tok/s
Qwen2.5-0.5B	—	512 / 128 tok	7,979 tok/s · 11 runs · range 7,661–7,993
Qwen2.5-0.5B	—	3968 / 128 tok	6,929 tok/s · 11 runs · range 6,856–6,952
Qwen2.5-0.5B	—	8192 / 128 tok	5,804 tok/s · 10 runs · range 5,519–5,822
Qwen2.5-0.5B	—	16384 / 128 tok	4,066 tok/s · 11 runs · range 3,848–4,363
Qwen2.5-0.5B	—	24576 / 128 tok	3,307 tok/s · 9 runs · range 3,039–3,514
Qwen2.5-0.5B	—	30000 / 128 tok	2,972 tok/s · 10 runs · range 2,825–3,094
Qwen2.5-1.5B	—	512 / 128 tok	2,296 tok/s · 11 runs · range 2,279–2,301
Qwen2.5-1.5B	—	3968 / 128 tok	2,188 tok/s · 9 runs · range 1,836–2,211
Qwen2.5-1.5B	—	8192 / 128 tok	1,824 tok/s · 11 runs · range 1,719–1,933
Qwen2.5-1.5B	—	16384 / 128 tok	1,506 tok/s · 11 runs · range 1,413–1,566
Qwen2.5-1.5B	—	24576 / 128 tok	1,266 tok/s · 11 runs · range 1,197–1,301
Qwen2.5-1.5B	—	30000 / 128 tok	1,169 tok/s · 7 runs · range 1,085–1,174
Qwen2.5-7B	—	512 / 128 tok	517.1 tok/s · 11 runs · range 450.9–525.1
Qwen2.5-7B	—	3968 / 128 tok	460.2 tok/s · 10 runs · range 434.9–480.4
Qwen2.5-7B	—	8192 / 128 tok	435.5 tok/s · 6 runs · range 405.3–463.0
Qwen2.5-7B	—	16384 / 128 tok	417.9 tok/s · 11 runs · range 406.1–423.8
Qwen2.5-7B	—	24576 / 128 tok	376.7 tok/s · 10 runs · range 358.6–377.9
Qwen3-1.7B	—	512 / 128 tok	2,161 tok/s · 11 runs · range 2,154–2,167
Qwen3-1.7B	—	3968 / 128 tok	1,969 tok/s · 10 runs · range 1,856–1,989
Qwen3-1.7B	—	8192 / 128 tok	1,629 tok/s · 11 runs · range 1,539–1,710
Qwen3-1.7B	—	16384 / 128 tok	1,346 tok/s · 10 runs · range 1,256–1,377
Qwen3-1.7B	—	24576 / 128 tok	1,133 tok/s · 11 runs · range 1,067–1,142
Qwen3-1.7B	—	30000 / 128 tok	1,020 tok/s · 10 runs · range 1,004–1,023

Every figure is veizik’s own measurement on this exact device, bound to the published build it was taken on. See Reproduce our benchmarks.

Compatibility

Architecture and its siblings.

Model family	Apple M4 Pro	State
Qwen2.5	Verified	Verified
Qwen3	Verified	Verified

No sibling device of this architecture is in the matrix.