Veizik Runtime sits between AI workloads and GPU hardware, coordinating execution, memory, precision, and hardware-specific runtime behavior through one interface — image and video models today, with language models on the roadmap as systems validation.
Make the model fit. Memory-aware execution expands the practical hardware envelope for modern video and image workloads.
Run through native paths. Hardware-specific runtime profiles select optimized execution paths for the target GPU.
Performance with measurable fidelity. Runtime optimization is paired with numerical and output-level validation — see how we measure.
One runtime across GPU tiers. Consumer, workstation, and datacenter configurations share one execution layer.
Video is the primary workload, image is second, and language models run as systems validation — proving runtime generality rather than serving as the core product today.
Wan 2.1/2.2, LTX-Video, HunyuanVideo†, CogVideoX, Step-Video — a native engine path for each, block-level numerically checked against reference.
primary workloadFLUX.1-dev, with SDXL/SD3.x-class support as models are qualified. Runs on the same runtime, same doctor scan, same entitlement.
primary workload† HunyuanVideo's licence excludes the EU, UK and South Korea — engine support exists, but Veizik (a Korean seller) does not publish outputs or figures for it. See model licences.
Compact consumer GPUs — LTX-Video-class workloads fit today.
Mainstream and professional desktop GPUs — the primary tier this runtime is measured on.
Workstation and datacenter configurations — larger models (Step-Video 30B class) and longer sequences.
Every engine path below is numerically checked at the block/engine level against a reference implementation before it ships.
| Model family | Engine path | Block-level rel_L2 | Status |
|---|---|---|---|
| LTX-Video | Native | 1.94e-7 | Verified |
| Wan 2.1 / 2.2 | Native | 5.53e-7 | Verified |
| CogVideoX | Native | 8.43e-7 | Verified |
| FLUX.1-dev (image) | Native | 4.58e-7 | Verified |
| Step-Video 30B | Native | 2.35e-6 | Verified |
| HunyuanVideo | Native | — | Licence-restricted in KR/EU/UK |
| Llama / Qwen (LLM) | Planned | — | Systems validation |
Numerical equivalence vs. the reference implementation at the block/engine level — not an end-to-end generation benchmark. Full evidence: Benchmarks.
Veizik operates at the execution layer, adapting supported workloads to GPU hardware through a unified runtime interface. Applications, interfaces, orchestration systems, and model pipelines connect to Veizik through a stable execution boundary — full technical detail in the runtime architecture doc.
Resume interrupted jobs and reuse validated prefixes where the selected runtime profile supports recovery — the good part of a render is kept, and only the failing tail regenerates. preview build
See the recovery benchmarkVeizik's core is an independent native execution layer — not a wrapper bolted onto someone else's stack. What's public is the surface you integrate against:
Independent execution engine, purpose-built per GPU architecture.
veizik doctor / login / status / t2v — see Developers.
Qualified model families plug into the runtime through a stable adapter interface.
NVIDIA, consumer through datacenter — see Hardware.
Run your existing graph under Veizik with low-VRAM + native engines — integration, not a dependency.
LLM endpoint compatibility for the systems-validation workload.