NVIDIA
Available now
Runs Language, Image and Video
- Ampere
- Ada Lovelace
- Hopper
- Blackwell
Independent AI execution engine
AI execution, re-engineered. One runtime across models, architectures and hardware — language, image and video, executed natively from the checkpoint you already have, with nothing else to install.
- NVIDIA Available now
- Apple Silicon Available now
- Edge Available now
Suggested for this device. Pick another platform any time.
bash · Linux x86-64
curl -fsSL https://veizik.com/install.sh | sh veizik install wan22 veizik activate veizik pull wan22 veizik run wan22_t2v_a14b "<your prompt>" --out ./clip # frames are written to ./clip/frames/
Output
zsh · macOS
# Apple Silicon · macOS (arm64) ./install.sh veizik run qwen2.5-1.5b-instruct "In two sentences, what is a safetensors checkpoint?" --tokens 96
Output
A safetensors checkpoint is a file that contains the weights and biases of a neural network model, which are used to train the model and generate the model's output. The safetensors format is designed to be more efficient and portable than the original TensorFlow checkpoint format, and it allows for better compression and storage of the model's weights. It is commonly used in the field of machine learning and artificial intelligence to save and load neural network models.
bash · Raspberry Pi
# Edge · Raspberry Pi curl -fsSL https://veizik.com/rpi/install.sh | sh veizik run smollm2_135m "The capital of France is" --mode balanced
Output
Paris
What do you want to run?
Start with the task. Each one opens the models that support it, with the devices and builds they run on.
Write and reason.
Chat, complete and serve language models on your own machine, with a local OpenAI-style endpoint when your tools need one.
See language models →Create images.
Turn a text prompt into a still image on your own GPU or edge board, from a published checkpoint.
See image models →Generate video.
Render text-to-video and image-to-video clips locally. Frames are written to a folder you choose.
See video models →See what your hardware can make.
Each result was rendered by VEIZIK from a published checkpoint. The caption names the model, the device and the build that made it.
A native runtime, on your terms.
Native execution.
Use a native runtime without a Python or deep-learning framework installation on supported deployment paths. VEIZIK compiles and runs the model itself, on the GPU or CPU you have.
Read about the product →Your checkpoints.
Start from supported published checkpoints. Weights come from their publishers and stay under their licences. Review each model’s requirements before you run it.
See supported models →Local generation.
Generation runs on your hardware, so prompts and results stay on your machine. Check the activation and download requirements before you begin.
Check activation →Start with the device you have.
Each platform shows what you can download today, read from the release registry. Pick the one you have and see the devices it covers.
Apple Silicon
Available now
Runs Language
- M1
- M2
- M3
- M4
Edge devices
Available now
Runs Language, Image, Video and Robotics control
- Raspberry Pi
- Android · Exynos
Measured on named hardware.
Our own numbers, each shown with its model, device, workload and build. Measured on Apple Silicon across 4 models and on NVIDIA GPUs across 5 models. Compare only runs that share a matching protocol.
Decode speed
197.9tok/s
Qwen3.6-35B-A3B on GeForce RTX 3090
- Single-token decoding · 22-token prompt · 128 tokens generated
- median of 3 runs, range 197.6–200.7 tok/s, measured 2026-10-22
- Build cuda-qwen-1.1.4. Carried from cuda-qwen-1.0.0. Measured on v1.0.0 of this package (2026-10-22), whose language GPU code is byte-identical to v1.1.0's; v1.1.0 was not re-timed.
GPU memory in use
21.44GiB
Qwen3.6-35B-A3B on GeForce RTX 3090
- Balanced mode
- single run, measured 2026-10-22
- Build cuda-qwen-1.1.4. Carried from cuda-qwen-1.0.0. Measured on v1.0.0 of this package (2026-10-22), whose language GPU code is byte-identical to v1.1.0's; v1.1.0 was not re-timed.
Decode speed
54.9tok/s
Qwen2.5-7B on Apple M4 Pro
- 512-token prompt · 128 tokens generated
- median of 11 runs, range 50.2–56.3 tok/s, measured 2026-10-04
- Build 2026.10.02-r10
Building AI into a product?
Explore VEIZIK for business: evaluate a compiler and runtime stack that controls model execution, memory and scheduling on supported hardware.
Explore VEIZIK for businessChoose your build.
Select your platform to see available packages, requirements and the next steps.
Basic is over. Veizik begins.
Powered by Wan-AI/Alibaba. Powered by Stability AI. Model weights are licensed by their publishers; VEIZIK authors the inference engine only and neither redistributes nor relicenses them.