VeizikNVIDIA and Apple Silicon builds are available now.Release notes →
veiziknative inference engine
Sign inDownload
Product
OverviewWhat the engine isHow it worksWeights in, native execution outPlatformsNVIDIA · Apple SiliconSecurity & licensingDevice-bound, local by designReleasesEvery build, dated
Engineering
All capabilitiesOne researched page per capabilityVideo memory720p that stays under a 24 GB cardA 35B model on one card21 GiB resident on a 24 GB GPUEdge optimizationGenerative AI on a Raspberry Pi CPUIts own stackNo Python, no framework, no cuBLASResume a renderCrash, resume, same result
Models
NVIDIALanguage and video — available nowApple SiliconQwen2.5 · Qwen3 — availableRaspberry Pi · edgeSD1.5 · SDXS · GPT-2 — arriving soonFull catalogueEvery model and its status
Hardware
All hardwareEvery device × every modelPick your deviceWhat runs on yoursNVIDIAAmpere → Blackwell · CUDAApple SiliconM1 – M4 · MetalRaspberry Pi · edgeCPU & NPU — generative AI on a Pi
Benchmarks
All resultsMatrix or pick, by buildBuilds & downloadsEvery build, JSON / CSVMethod & notationHow every figure is taken
Docs
QuickstartInstall to first runCommand referenceEvery documented commandActivate a machineDevices and licenceWhat leaves the machineOffline and networkTroubleshootingWhen something fails
GalleryReleasesSign inDownloadGitHub
← Models
Language · Apple Silicon · Metal

Qwen2.5

Run Qwen2.5 directly from supported safetensors checkpoints with no converted model copy.

Status
Available
Platform
Apple Silicon · Metal
Class
Language
Publisher
Alibaba (Qwen)
Architecture
Qwen2
Weights
Supplied by you
Guide

How to run it

  1. Install

    One signed package for Apple Silicon. Nothing else to install first.

    step 1
    $ ./install.sh
  2. Activate this device

    Approve the machine in your browser; it receives a licence bound to it.

    step 2
    $ veizik activate
  3. Pull the checkpoint

    Downloads the published safetensors into the local cache. No conversion.

    step 3
    $ veizik pull qwen2.5-0.5b
  4. Run it

    One-shot, interactive, or as a local HTTP endpoint.

    step 4
    $ veizik run qwen2.5-0.5b "The capital of France is"
$ veizik chat qwen2.5-0.5b
$ veizik serve qwen2.5-0.5b --port 8080
Support

Where it runs.

Evidence and availability are separate: a verified device is not a download, and an available model is not verified on every device.

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.
FamilyRTX 3080 TiRTX 3090RTX 4090RTX 5090RTX A5000A40A100 80GB PCIeA100-SXM4-80GBH100 80GB HBM3RTX PRO 4500 BlackwellM4 ProM1 UltraM1 MaxM2 Pro
Qwen2.5
all variants		Not offeredNot offeredNot offeredNot offeredNot offeredNot offeredNot offeredNot offeredNot offeredNot offeredVerifiedVerifiedVerifiedVerified

Availability by variant

Model idAvailabilityPlatformContextWeights in memory
qwen2.5-0.5bAvailableApple Silicon32,768343.6 MB
measured, deterministic load
qwen2.5-1.5bAvailableApple Silicon32,768926.9 MB
measured, deterministic load
qwen2.5-7bAvailableApple Silicon32,7684,114.0 MB
measured, deterministic load
qwen2.5-0.5b-instructNot publishedApple Silicon——
qwen2.5-1.5b-instructNot publishedApple Silicon——
qwen2.5-7b-instructNot publishedApple Silicon——

Verified on

  • Apple M1 Max · build 2026.10.02-r10 · 2026-10-20 · published measurement
  • Apple M1 Ultra · build 2026.10.02-r10 · 2026-10-09 · published measurement
  • Apple M2 Pro · build 2026.10.02-r10 · 2026-10-06 · published measurement
  • Apple M4 Pro · build 2026.10.02-r10 · 2026-10-07 · published measurement

“Verified” means this family ran end to end on that exact device, on a published build. Architecture membership never upgrades a device to verified.

Variants

6 checkpoints

Model idSizeCheckpointOn diskContextWeights in memory
qwen2.5-0.5b0.5BQwen/Qwen2.5-0.5B942 MiB32,768343.6 MB
qwen2.5-1.5b1.5BQwen/Qwen2.5-1.5B2,944 MiB32,768926.9 MB
qwen2.5-7b7BQwen/Qwen2.5-7B14,525 MiB32,7684,114.0 MB
qwen2.5-0.5b-instruct0.5BQwen/Qwen2.5-0.5B-Instruct———
qwen2.5-1.5b-instruct1.5BQwen/Qwen2.5-1.5B-Instruct———
qwen2.5-7b-instruct7BQwen/Qwen2.5-7B-Instruct———
Performance

Measured on the released build

ModelDeviceDecode tok/sPrefill tok/sRuns
Qwen2.5-0.5BApple M1 Ultra377.49,63211
Qwen2.5-0.5BApple M2 Pro300.65,06011
Qwen2.5-0.5BApple M1 Max273.17,90111
Qwen2.5-0.5BApple M4 Pro462.87,97911
Qwen2.5-1.5BApple M1 Ultra260.24,03211
Qwen2.5-1.5BApple M2 Pro144.61,50611
Qwen2.5-1.5BApple M1 Max181.12,44011
Qwen2.5-1.5BApple M4 Pro215.02,29611
Qwen2.5-7BApple M1 Ultra104.71,00911
Qwen2.5-7BApple M2 Pro38.3333.111
Qwen2.5-7BApple M1 Max60.4545.911
Qwen2.5-7BApple M4 Pro54.9517.111

512-token prompt, median of runs, lower bound. Every prompt length and device →

Licence

Weights and terms

Qwen2.5 weights are published by Alibaba under their own license. Veizik neither redistributes nor relicenses them.

Full licence terms →

Other models linked in the sidebar:

Stable Diffusion 3.5: Powered by Stability AI HunyuanImage: Weights are not included. Model availability and use are governed by the model provider's license and regional restrictions.