veizik-cuda
Run language, video and image models on your own NVIDIA GPU with one native engine.
Veizik runs Qwen3.5 / Qwen3.6 language models and Wan2.1 / Wan2.2 video models on Linux with an NVIDIA GPU. It reads the Hugging Face checkpoint you downloaded yourself. No Python, no framework, no CUDA toolkit, no cloud at inference time: only the NVIDIA driver. Prompts, generated text and generated frames stay on the machine.
Who this is for
Developers and teams who run open models on their own NVIDIA hardware and want a sealed native engine instead of a framework stack: a one-shot run, an interactive chat, or a local HTTP server (serve) — an OpenAI-compatible API for language models and a render API for video — with the same commands for every model.
What it looks like
$ veizik run qwen35_2b "The capital of France is" --tokens 32
$ veizik chat qwen35_2b
$ veizik serve qwen35_2b --port 8000 # OpenAI API: POST /v1/chat/completions
$ veizik run wan21_t2v_1_3b "A red car driving along a coastal road at sunset." --out $HOME/clip
$ ls $HOME/clip/frames # frame_00000.png ...
$ veizik serve wan21_t2v_1_3b --port 8000 # render API: POST /v1/videos/generations
Supported environment and models
- Linux x86-64, running natively, with an NVIDIA GPU of compute capability 8.0 or newer and a driver that provides
libcuda.so.1. No CUDA toolkit and no Python are needed. WSL is not a supported runtime for the sealed engines. One GPU runs one model.
- GPU code ships as a separate precompiled set per architecture:
sm_80 (A100), sm_86 (RTX 3090, RTX A5000, A40), sm_89 (RTX 4090, L4, L40S), sm_90a (H100), sm_100a (B200), sm_120a (RTX 5090, RTX PRO 6000 Blackwell). A GPU architecture not in this list has no GPU code in these packages.
- Every device is activated once (https://veizik.com/activate). The engines open their GPU code only on an activated device.
- Model weights are not included. You download each checkpoint from its publisher, under the publisher's licence.
Language — veizik-qwen-cuda 1.1.3 (5 models in one package)
| Model key | Model | Hugging Face checkpoint | Chat | Image and video input | Resident GPU memory, balanced |
|---|
qwen35_0p8b | Qwen3.5-0.8B | Qwen/Qwen3.5-0.8B | yes | — | ≈ 2.5 GiB |
qwen35_2b | Qwen3.5-2B | Qwen/Qwen3.5-2B | yes | — | ≈ 3.7 GiB |
qwen35_4b | Qwen3.5-4B | Qwen/Qwen3.5-4B | yes | — | ≈ 5.7 GiB |
qwen35_9b | Qwen3.5-9B | Qwen/Qwen3.5-9B | yes | — | ≈ 8.7 GiB |
qwen36_35b_a3b | Qwen3.6-35B-A3B (mixture of experts, about 3B active per token) | Qwen/Qwen3.6-35B-A3B | yes | yes — --image / --video | ≈ 21.7 GiB (text) |
All five models are marked EXPERIMENTAL in the signed model catalogue. Memory is measured after load with text input on an RTX 5090, on the 1.0.0 build; long prompts, images, video and serving need more. Image and video input is an option of the package's launcher for Qwen3.6-35B-A3B only: one PNG or JPEG image, or one folder of 2–256 PNG or JPEG frames, per run, together with a text prompt. The image and video stage ships GPU code for sm_86, sm_89 and sm_120a. On a GPU the image and video stage does not support, the run is refused cleanly with code 50 (the image stage refused this request (status 50)) before any image or video work. Validation ran on RTX 3090, RTX A5000, RTX 4090 and RTX 5090; the sm_80, sm_90a and sm_100a GPU code is included but was not run in validation.
Video — veizik-wan 1.2.8 (4 models in one package, with a resident render server and mid-render preempt)
| Model key | Model | Hugging Face checkpoint | Task | Defaults (size · frames · steps) |
|---|
wan21_t2v_1_3b | Wan2.1-T2V-1.3B | Wan-AI/Wan2.1-T2V-1.3B-Diffusers | text to video | 832×480 · 49 · 30 |
wan21_t2v_14b | Wan2.1-T2V-14B | Wan-AI/Wan2.1-T2V-14B-Diffusers | text to video | 832×480 · 49 · 30 |
wan22_t2v_a14b | Wan2.2-T2V-A14B (two experts) | Wan-AI/Wan2.2-T2V-A14B-Diffusers | text to video | 1280×720 · 81 · 40 |
wan22_i2v_a14b | Wan2.2-I2V-A14B (two experts) | Wan-AI/Wan2.2-I2V-A14B-Diffusers | image to video | 832×480 · 81 · 40 |
The WAN package carries six kernel sets, one per architecture: sm_80, sm_86, sm_89, sm_90a, sm_100a and sm_120a. Every speed figure Veizik states for this bundle was measured on an RTX 3090 (sm_86) only; other GPUs run their own kernel sets at speeds not measured.
Install and first run
Step 1 installs the veizik command (the latest release); it then installs the packages. On macOS the script points you to the Apple disk image instead.
curl -fsSL https://veizik.com/install.sh | sh # installs the veizik command only, sha-checked against this repository's SHA256SUMS
export PATH="$HOME/.local/bin:$PATH"
veizik list # model engines: installed and available
veizik install qwen36 # Qwen package (release v1.1.3)
veizik install wan22 # WAN package (release v1.2.8)
veizik activate # opens veizik.com/activate — approve this machine once
# no browser (SSH, servers)? create a key under "Browser-free activation" and a password under "CLI password"
# on your account page, then: veizik activate --key vzek_… --password-stdin < pwfile
export VZ_WEIGHTS=/data/models # a folder holding Qwen3.5-0.8B/ and the other checkpoints you downloaded
veizik run qwen35_2b "The capital of France is" --tokens 32
veizik run qwen36_35b_a3b "Explain attention in one paragraph."
veizik chat qwen35_2b # interactive REPL (needs an activated device + weights)
veizik serve qwen35_2b --port 8000 # OpenAI API: POST /v1/chat/completions (needs activation)
veizik run wan22_t2v_a14b "A red car driving along a coastal road at sunset." --out $HOME/clip
veizik serve wan22_t2v_a14b --port 8000 # WAN render API: POST /v1/videos/generations (needs activation)
veizik ps # resident servers running here
veizik stop 8000 # stop the server on a port (or: veizik stop all)
veizik uninstall qwen36 # removes an installed package (alias: veizik rm)
- Every model answers the same commands:
run, chat, serve, ps, stop, models, list, install, update, uninstall / rm, pull, hf, remove, convert, inspect, bench, cache, config, completion, status, activate, devices, device-id, doctor and version; veizik <command> --help gives the detail and cuda_deploy/CLI_USAGE.md is the full reference. veizik update and veizik convert exist but are not functional yet.
- Install names are keywords: the Qwen package is
qwen, qwen35, qwen36 or any of its model keys (qwen35_0p8b, qwen35_2b, qwen35_4b, qwen35_9b, qwen36_35b_a3b); the WAN package is wan, wan21, wan22 or any of its model keys (wan21_t2v_1_3b, wan21_t2v_14b, wan22_t2v_a14b, wan22_i2v_a14b). Case does not matter, and . or - may stand for _.
veizik install downloads the package anonymously from this repository's release, refuses it if its sha256 differs from the published release digest, and installs it to ${XDG_STATE_HOME:-~/.local/state}/veizik/engines/<model>/. It refuses to run on macOS. veizik run, veizik chat and veizik serve find the installed engine themselves.veizik chat is an interactive REPL; veizik serve hosts a local HTTP server on 127.0.0.1 (omit the model to serve every installed one). For Qwen, serve is OpenAI-compatible (POST /v1/chat/completions, /v1/completions, /v1/models, health endpoints; --require-auth --keyfile requires an API key); for WAN, is a render API ( with → a folder of PNG frames; , ). , and generate only after with the model weights present (, where the command takes it, or the cache); until the device is activated a request is refused (HTTP 403 / exit 70).
Checksums (published release digests)
veizik install downloads each package and verifies it against its published release digest automatically — there is no separate manual install step. For reference (e.g. to verify a direct download), the published downloads and digests are:
What needs the network: installing a package, creating the account, activating the device and renewing its lease. Generation itself does not.
Measured on this build
- Qwen: the published rows and the rows committed in
cuda_deploy/bench/ were measured on 1.0.0 builds of this package (2026-10-21), with text input. They were not re-measured on 1.1.0, and the board marks every row measured on an earlier build. No figure is published for image or video input. Board: https://veizik.com/benchmarks
- Qwen output: greedy decoding is the default and is deterministic. For the committed benchmark prompt the output hash was identical on the RTX 3090, RTX 4090 and RTX 5090 in every measured cell. Sampling (
--temperature, --top-p, --top-k, --seed) is experimental.
- WAN: the same inputs give the same latent bytes on one GPU architecture. Reference hashes for each model are in the bundle's
CONFORMANCE.json (measured on an RTX 3090); decoded frames are byte-exact with --det-vae yes.
Full benchmarks and how to reproduce
https://veizik.com/benchmarks — every published row names the GPU, model, mode, operating point, run count and spread. Rows measured on an earlier build are marked. The Qwen rows, the per-run lines they were computed from, and a kit that re-measures them on your GPU through the installed CLI are in cuda_deploy/bench/ — start at REPRODUCIBILITY.md. The WAN bundle's REPRO.md describes how to check a render against CONFORMANCE.json.
Limits, troubleshooting, terms
- Qwen: Qwen3.6-35B-A3B in
quality mode needs about 26 GB or more and did not complete on a 32 GB card in validation; use balanced or fast. Image and video input is available for Qwen3.6-35B-A3B only (veizik run … --image / --video), one image or one video per run. Not in this release: LoRA adapters, audio input, tool calling, embeddings and logprobs over HTTP, and a weight converter (veizik convert exists but does nothing useful yet; so does veizik update).
- WAN: a run that is interrupted continues with
veizik run … --resume and the same --out (or --resume-from REF), bit-identical to an uninterrupted render.
- WAN serve:
veizik serve exposes a render API (POST /v1/videos/generations with a prompt and seed → a folder of PNG frames; /health, /v1/models) on a resident render server (bin/wan_res). One process and one GPU context render request after request; results are bit-identical to a one-shot run; every request re-checks the licence, and a request can be preempted at a step boundary — the engine checkpoints and yields the GPU, and the preempted request then resumes from its checkpoint and continues to a bit-identical result. It needs veizik activate and the weights, like run. See cuda_deploy/wan/README.md §8.
- WAN provenance: every generated frame carries a v1 provenance mark — an in-pixel watermark plus XMP
DigitalSourceType=trainedAlgorithmicMedia. The mark holds no prompt, device, path or account data.
- Integrity: both engines check their own files at every start. A changed, added or missing file stops the run with exit code 4; reinstall from a verified package. Do not add files inside the installed or unpacked engine folder; write output elsewhere.
- Licensing: a device runs only with a valid lease. Exit 70 means activation is required (
veizik activate); exit 74 means the lease present was not issued for this package on this device, or does not enable what was asked.
- Not supported: Windows, macOS, WSL, GPUs below compute capability 8.0, and splitting one model across several GPUs. The Qwen package also contains an aarch64
veizik binary, but the engine is x86-64 only.
- Troubleshooting:
veizik doctor · veizik status · §6 · · · Issues in this repository ·
Package contents
veizik-qwen-cuda-1.1.3-linux-x86_64.vzpack (v1.1.3, sha256 0fcee30f…, 56,552,565 B)
A single-file .vzpack (content-addressed, with a signed .vzindex) that veizik install qwen verifies against the published digest and materializes to the engine folder veizik-qwen-cuda-0.1.0-rc1-linux-x86_64-sm86/: README.txt 1,927 B · install.sh 3,882 B · SHA256SUMS 207,093 B · veizik-x86_64 1,534,320 B · veizik-aarch64 994,728 B · PTXSEAL.prov 124 B · PTXSEAL.prov.sig 64 B · the sealed engine folder of the same name: INTERFACE.json 10,540 B · PIPELINE.sh 108 B · SHA256SUMS.vz 132,541 B · models.json 644 B · models.json.sig 64 B · bin/ (8 files, including the image and video runners) · models/ (5 model folders, per-model metadata) · GPU code, 251 files per architecture: ptx/ (sm_80) · ptx_sm86/ · ptx_sm89/ · ptx_sm90a/ · ptx_sm100a/ · ptx_sm120a/ · image and video stage GPU code, 16 files per architecture: ptx_vis_sm86/ · ptx_vis_sm89/ · ptx_vis_sm120a/. 1,583 files in all. Verify with sha256sum -c SHA256SUMS inside the top folder.
veizik-wan-1.2.8-linux-x86_64.vzpack (v1.2.8, sha256 3fbe53c3…, 6,851,149 B)
A single-file .vzpack (content-addressed, signed .vzindex, shared-core veizik-shared-core.vzroot) that veizik install wan22 verifies against the published digest, checks its signed index and materializes to the engine folder. Same sealed engine and all six per-architecture GPU kernel sets (identical engine; all-architecture support).
Engine folder (materialized from the .vzpack by veizik install wan22)
PTXSEAL.prov 124 B · PTXSEAL.prov.sig 64 B · folder veizik-wan-1.2.0-linux-x86_64/: veizik 4,329 B · PIPELINE.sh 48,411 B · README.txt 8,307 B · MODEL.txt 418 B · STAGES.txt 1,239 B · INTERFACE.json 22,444 B · manifest.json 416 B · CONFORMANCE.json 4,099 B · REPRO.md 670 B · PROVENANCE.txt 483 B · THIRD_PARTY_NOTICES.txt 1,070 B · SHA256SUMS.vz 14,200 B · inject.sh 3,390 B · s6_segments.sh 998 B · vz_modes.conf 312 B · bin/ (13 files, incl. the new veizik-resident resident host-loop) · bin_seam/ (1 file) · lib/ (15 libraries) · models/ (model folders) · GPU kernel sets: ptx/ 17 · ptx_89/ 17 · ptx_a80/ 18 · ptx_h90/ 22 · ptx_b100a/ 20 · ptx_b120/ 17 files. 173 files in all. The engine, opener and all six GPU kernel sets are byte-identical to 1.2.2/1.2.3; bin/wan_s2 is the 1.2.4 fix for CUDA-13.x non-root rendering. The folder keeps the version-stable name veizik-wan-1.2.0-linux-x86_64 so veizik activate resolves the same release id across versions. Verify with bin/vzseal <folder> (prints VZSEAL OK 169).
