VeizikNVIDIA, Apple Silicon and edge builds are available now.Release notes →
veiziknative inference engine
For business Sign inDownload
Hardware

Apple M1 Ultra

Veizik on Apple M1 Ultra — Apple Silicon · Metal · M1.

Overview

What this device is.

Apple M1 UltraAvailable
Vendor
Apple Silicon
Backend
Metal
Platform
macOS on Apple Silicon (arm64)
Generation
M1
Architecture
M1
Availability
Available

On Apple M1 Ultra, 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
M1
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 familyEvidence on this deviceAvailability
Qwen2.5
Language		Verified VerifiedAvailable
Qwen3
Language		Verified VerifiedAvailable
Performance

Measured on this exact device.

ModelTaskWeights in memory
Qwen2.5-0.5BLanguage343.6 MB
Qwen2.5-1.5BLanguage926.9 MB
Qwen2.5-7BLanguage4,114.0 MB
Qwen3-1.7BLanguage968.0 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).

ModelModeWorkloadtok/s
Qwen2.5-0.5B—512 / 128 tok377.4 tok/s · 11 runs · range 355.8–385.6
Qwen2.5-0.5B—512 / 128 tok377.4 tok/s · 11 runs · range 355.8–385.6
Qwen2.5-0.5B—3968 / 128 tok305.5 tok/s · 11 runs · range 304.4–313.3
Qwen2.5-0.5B—3968 / 128 tok305.5 tok/s · 11 runs · range 304.4–313.3
Qwen2.5-0.5B—8192 / 128 tok280.6 tok/s · 11 runs · range 278.7–286.6
Qwen2.5-0.5B—8192 / 128 tok280.6 tok/s · 11 runs · range 278.7–286.6
Qwen2.5-0.5B—16384 / 128 tok259.1 tok/s · 11 runs · range 257.4–262.0
Qwen2.5-0.5B—16384 / 128 tok259.1 tok/s · 11 runs · range 257.4–262.0
Qwen2.5-0.5B—24576 / 128 tok205.6 tok/s · 11 runs · range 204.4–207.8
Qwen2.5-0.5B—24576 / 128 tok205.6 tok/s · 11 runs · range 204.4–207.8
Qwen2.5-0.5B—30000 / 128 tok199.0 tok/s · 11 runs · range 196.9–200.3
Qwen2.5-0.5B—30000 / 128 tok199.0 tok/s · 11 runs · range 196.9–200.3
Qwen2.5-1.5B—512 / 128 tok260.2 tok/s · 11 runs · range 244.7–261.2
Qwen2.5-1.5B—512 / 128 tok260.2 tok/s · 11 runs · range 244.7–261.2
Qwen2.5-1.5B—3968 / 128 tok233.9 tok/s · 11 runs · range 215.3–241.2
Qwen2.5-1.5B—3968 / 128 tok233.9 tok/s · 11 runs · range 215.3–241.2
Qwen2.5-1.5B—8192 / 128 tok201.0 tok/s · 11 runs · range 200.2–204.0
Qwen2.5-1.5B—8192 / 128 tok201.0 tok/s · 11 runs · range 200.2–204.0
Qwen2.5-1.5B—16384 / 128 tok169.1 tok/s · 11 runs · range 168.1–170.5
Qwen2.5-1.5B—16384 / 128 tok169.1 tok/s · 11 runs · range 168.1–170.5
Qwen2.5-1.5B—24576 / 128 tok145.8 tok/s · 11 runs · range 144.8–147.6
Qwen2.5-1.5B—24576 / 128 tok145.8 tok/s · 11 runs · range 144.8–147.6
Qwen2.5-1.5B—30000 / 128 tok142.6 tok/s · 11 runs · range 141.7–143.7
Qwen2.5-1.5B—30000 / 128 tok142.6 tok/s · 11 runs · range 141.7–143.7
Qwen2.5-7B—512 / 128 tok104.7 tok/s · 11 runs · range 104.0–105.1
Qwen2.5-7B—512 / 128 tok104.7 tok/s · 11 runs · range 104.0–105.1
Qwen2.5-7B—3968 / 128 tok88.2 tok/s · 11 runs · range 87.8–88.5
Qwen2.5-7B—3968 / 128 tok88.2 tok/s · 11 runs · range 87.8–88.5
Qwen2.5-7B—8192 / 128 tok84.0 tok/s · 11 runs · range 83.6–84.3
Qwen2.5-7B—16384 / 128 tok72.8 tok/s · 11 runs · range 72.5–73.0
Qwen2.5-7B—16384 / 128 tok72.8 tok/s · 11 runs · range 72.5–73.0
Qwen2.5-7B—24576 / 128 tok65.7 tok/s · 11 runs · range 65.5–65.8
Qwen2.5-7B—24576 / 128 tok65.7 tok/s · 11 runs · range 65.5–65.8
Qwen2.5-7B—30000 / 128 tok60.9 tok/s · 11 runs · range 60.8–61.4
Qwen2.5-7B—30000 / 128 tok60.9 tok/s · 11 runs · range 60.8–61.4
Qwen3-1.7B—512 / 128 tok289.8 tok/s · 11 runs · range 275.2–292.0
Qwen3-1.7B—512 / 128 tok289.8 tok/s · 11 runs · range 275.2–292.0
Qwen3-1.7B—3968 / 128 tok252.8 tok/s · 11 runs · range 238.6–253.4
Qwen3-1.7B—3968 / 128 tok252.8 tok/s · 11 runs · range 238.6–253.4
Qwen3-1.7B—8192 / 128 tok200.7 tok/s · 11 runs · range 199.9–202.2
Qwen3-1.7B—8192 / 128 tok200.7 tok/s · 11 runs · range 199.9–202.2
Qwen3-1.7B—16384 / 128 tok155.7 tok/s · 11 runs · range 155.2–157.7
Qwen3-1.7B—16384 / 128 tok155.7 tok/s · 11 runs · range 155.2–157.7
Qwen3-1.7B—24576 / 128 tok127.0 tok/s · 11 runs · range 126.1–127.3
Qwen3-1.7B—24576 / 128 tok127.0 tok/s · 11 runs · range 126.1–127.3
Qwen3-1.7B—30000 / 128 tok114.3 tok/s · 11 runs · range 113.8–114.6
Qwen3-1.7B—30000 / 128 tok114.3 tok/s · 11 runs · range 113.8–114.6

Prefill, tokens per second Measured here

Prompt tokens processed per second.

ModelModeWorkloadtok/s
Qwen2.5-0.5B—512 / 128 tok9,632 tok/s · 11 runs · range 9,088–10,741
Qwen2.5-0.5B—512 / 128 tok9,632 tok/s · 11 runs · range 9,088–10,741
Qwen2.5-0.5B—3968 / 128 tok13,075 tok/s · 11 runs · range 12,868–13,274
Qwen2.5-0.5B—3968 / 128 tok13,075 tok/s · 11 runs · range 12,868–13,274
Qwen2.5-0.5B—8192 / 128 tok11,328 tok/s · 11 runs · range 11,288–11,399
Qwen2.5-0.5B—8192 / 128 tok11,328 tok/s · 11 runs · range 11,288–11,399
Qwen2.5-0.5B—16384 / 128 tok8,509 tok/s · 11 runs · range 8,488–8,526
Qwen2.5-0.5B—16384 / 128 tok8,509 tok/s · 11 runs · range 8,488–8,526
Qwen2.5-0.5B—24576 / 128 tok6,773 tok/s · 11 runs · range 6,765–6,777
Qwen2.5-0.5B—24576 / 128 tok6,773 tok/s · 11 runs · range 6,765–6,777
Qwen2.5-0.5B—30000 / 128 tok5,895 tok/s · 11 runs · range 5,892–5,899
Qwen2.5-1.5B—512 / 128 tok4,032 tok/s · 11 runs · range 3,945–4,039
Qwen2.5-1.5B—512 / 128 tok4,032 tok/s · 11 runs · range 3,945–4,039
Qwen2.5-1.5B—3968 / 128 tok4,470 tok/s · 11 runs · range 4,325–4,478
Qwen2.5-1.5B—3968 / 128 tok4,470 tok/s · 11 runs · range 4,325–4,478
Qwen2.5-1.5B—8192 / 128 tok4,021 tok/s · 11 runs · range 3,975–4,043
Qwen2.5-1.5B—8192 / 128 tok4,021 tok/s · 11 runs · range 3,975–4,043
Qwen2.5-1.5B—16384 / 128 tok3,309 tok/s · 11 runs · range 3,302–3,314
Qwen2.5-1.5B—16384 / 128 tok3,309 tok/s · 11 runs · range 3,302–3,314
Qwen2.5-1.5B—24576 / 128 tok2,796 tok/s · 11 runs · range 2,738–2,801
Qwen2.5-1.5B—24576 / 128 tok2,796 tok/s · 11 runs · range 2,738–2,801
Qwen2.5-1.5B—30000 / 128 tok2,518 tok/s · 11 runs · range 2,481–2,524
Qwen2.5-1.5B—30000 / 128 tok2,518 tok/s · 11 runs · range 2,481–2,524
Qwen2.5-7B—512 / 128 tok1,009 tok/s · 11 runs · range 1,007–1,010
Qwen2.5-7B—512 / 128 tok1,009 tok/s · 11 runs · range 1,007–1,010
Qwen2.5-7B—3968 / 128 tok1,038 tok/s · 11 runs · range 1,035–1,039
Qwen2.5-7B—3968 / 128 tok1,038 tok/s · 11 runs · range 1,035–1,039
Qwen2.5-7B—8192 / 128 tok975.9 tok/s · 11 runs · range 965.0–976.2
Qwen2.5-7B—8192 / 128 tok975.9 tok/s · 11 runs · range 965.0–976.2
Qwen2.5-7B—16384 / 128 tok865.2 tok/s · 11 runs · range 852.8–865.7
Qwen2.5-7B—16384 / 128 tok865.2 tok/s · 11 runs · range 852.8–865.7
Qwen2.5-7B—24576 / 128 tok769.2 tok/s · 11 runs · range 763.3–776.2
Qwen2.5-7B—24576 / 128 tok769.2 tok/s · 11 runs · range 763.3–776.2
Qwen2.5-7B—30000 / 128 tok720.1 tok/s · 11 runs · range 715.2–723.3
Qwen2.5-7B—30000 / 128 tok720.1 tok/s · 11 runs · range 715.2–723.3
Qwen3-1.7B—512 / 128 tok4,213 tok/s · 11 runs · range 3,580–4,248
Qwen3-1.7B—512 / 128 tok4,213 tok/s · 11 runs · range 3,580–4,248
Qwen3-1.7B—3968 / 128 tok4,171 tok/s · 11 runs · range 4,095–4,183
Qwen3-1.7B—3968 / 128 tok4,171 tok/s · 11 runs · range 4,095–4,183
Qwen3-1.7B—8192 / 128 tok3,732 tok/s · 11 runs · range 3,728–3,746
Qwen3-1.7B—8192 / 128 tok3,732 tok/s · 11 runs · range 3,728–3,746
Qwen3-1.7B—16384 / 128 tok3,001 tok/s · 11 runs · range 2,968–3,004
Qwen3-1.7B—16384 / 128 tok3,001 tok/s · 11 runs · range 2,968–3,004
Qwen3-1.7B—24576 / 128 tok2,503 tok/s · 11 runs · range 2,491–2,503
Qwen3-1.7B—24576 / 128 tok2,503 tok/s · 11 runs · range 2,491–2,503
Qwen3-1.7B—30000 / 128 tok2,235 tok/s · 11 runs · range 2,229–2,237
Qwen3-1.7B—30000 / 128 tok2,235 tok/s · 11 runs · range 2,229–2,237

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.

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 familyApple M1 UltraState
Qwen2.5VerifiedVerified
Qwen3VerifiedVerified

Same architecture (M1), so a family verified on one carries to the others as Supported: Apple M1 Max.