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
How it worksOne researched page per capabilityVideo memory720p that stays under a 24 GB cardA 35B model on one card21 GiB resident on a 24 GB GPUIts own stackNo Python, no framework, no cuBLASResume a renderCrash, resume, same result
Models
NVIDIALanguage and video — available nowApple SiliconQwen2.5 · Qwen3 — availableFull catalogueEvery model and its status
Hardware
All hardwareEvery device × every modelPick your deviceWhat runs on yoursNVIDIAAmpere → Blackwell · CUDAApple SiliconM1 – M4 · Metal
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
ReleasesSign inDownloadGitHub
Releases

veizik Qwen (CUDA) v1.1.2 — Qwen3.5 / Qwen3.6, multimodal — cuda-qwen-1.1.2

Qwen3.5 (0.8B, 2B, 4B, 9B) and Qwen3.6-35B-A3B on your own NVIDIA GPU, on-device, now with image and video input for Qwen3.6-35B-A3B. Veizik runs the models through its own native CUDA engine: no Python runtime, no CUDA toolkit, no cloud at inference time. You supply the weights; nothing leaves the machine.

This page records the state as of 2026-10-27 and is not updated as the catalogue moves. For what runs now, see models.

Highlights

What this build does

  • Image and video input in the package's Qwen3.6-35B-A3B engine: it reads one image (PNG or JPEG) or one video (2 to 256 frames as PNG or JPEG files, in file-name order) together with a text prompt. Pass --image FILE or --video DIR to veizik run qwen36_35b_a3b.
  • veizik serve <model>: a resident, batched, OpenAI-compatible HTTP server on 127.0.0.1 with server-sent-event streaming (/v1/chat/completions, /v1/completions, /v1/models, plus health and metrics endpoints).
  • API-key authentication for the server: --require-auth with --keyfile or --api-key; a request without the key answers HTTP 401.
  • veizik run for the five models with five modes (quality, balanced, fast, ultrafast, low-memory); the prompt can come from text or a file.
  • Native GPU code for sm_80 through sm_120a, with the integrity of the bundle checked at every start.
  • Device-bound activation, per device rather than per model: browser approval, or browser-free activation with an account key and CLI password.
Support

Models in this release

ModelTaskContextStatus
Qwen3.5-0.8BLanguage4,096Available
Qwen3.5-2BLanguage4,096Available
Qwen3.5-4BLanguage4,096Available
Qwen3.5-9BLanguage4,096Available
Qwen3.6-35B-A3BLanguage4,096Available
Platform
Linux on x86-64
Verification
This build is superseded by cuda-qwen-1.1.3. Install the current build with `veizik install qwen`.
File
a previous build — not downloadable
Size
59,008,512 bytes
SHA-256
d1f5fd35c950652a31dd70b6dc9baaef67e700f299033e8a83d7f0ab0e6184e6
Installation

Verify, install, check

zsh
$ shasum -a 256 -c SHA256SUMS
$ ./install.sh
$ veizik doctor

Full steps are in the documentation.

Fixed

Fixed issues

  • Serve no longer fails to boot on the tokenizer environment (serve GE_BOOT). Renders and language output are byte-identical to 1.1.1 (engine and GPU code unchanged).
  • veizik serve now starts and listens — a startup integrity check that could stall the local server in 1.1.0 is resolved; text and OpenAI-compatible serving work.
  • veizik run now returns decoded, readable text — 1.1.0 could print raw token ids. The token stream is available only with an explicit option.
Known issues

Known issues in cuda-qwen-1.1.2

  • Image and video input is a capability of the Qwen3.6-35B-A3B engine only, one image or one video per run, and its GPU code ships for RTX 30-series (sm_86), RTX 40-series (sm_89) and RTX 50-series (sm_120a) class GPUs only. The text models keep all six GPU architectures.
  • The server is loopback-only; to serve other machines put a TLS reverse proxy in front of it. Without --require-auth any local process can use it.
  • No usage ledger or quotas in the server; --usage-log, --quota-* and --max-tokens are accepted but not applied.
  • veizik stop does not see the server; stop a detached server by signalling the engine process.
  • veizik chat gives an interactive text session; veizik serve (/v1/chat/completions) is the validated path. Generation needs the model weights present and the device activated with veizik activate.
  • Qwen3.6-35B-A3B in quality mode needs about 26 GB or more and did not complete on a 32 GB RTX 5090 in validation; use balanced or fast.
  • Sampling options on run (--seed, --temperature, --top-p, --top-k) are opt-in and experimental; greedy decoding is the default and the verified path.
  • The server context window is 4,096 tokens per request (prompt plus generated) in the validated configuration.
  • GPU code for sm_80, sm_90a and sm_100a is included but was not run in validation.
  • WSL is not supported; the package installs on Linux x86-64 only.
  • Not in this release: LoRA adapters, image or video input on models other than Qwen3.6-35B-A3B, tool or function calling, logprobs and embeddings over HTTP, a weight converter.