Skip to content

VulRunWG · The runtime

Local inference.
Compute beyond AI.

A Vulkan-based GPU runtime with its own inference kernels and a path for custom WGSL workloads.

Run supported language, image and speech models. Give particles and flocking their own compute kernels. VulRunWG supplies shared GPU infrastructure beneath different applications and workloads.

Proprietary runtime · Public release preparation in progress

VulRunWGExecution architecture

From an idea to GPU work.

Local models

Text · Images · Speech

Supported model families
Custom compute

Particles · Flocking

Workload-specific WGSL
VulRunWGGPU kernels · Buffers · Command submission
wgpu / VulkanCompatible GPU + driver
Shared GPU infrastructure, distinct kernels for each workload. Hardware features determine the available execution paths.

What makes it different

The engine does the work.

Purpose-built GPU kernels

The local inference path implements model operations through its own GPU kernels, including matrix operations, attention and quantized-weight processing. It is not a wrapper around llama.cpp.

A Vulkan execution path

WGSL compute runs through wgpu and Vulkan on compatible graphics hardware. Recorded development runs cover NVIDIA and AMD configurations without requiring CUDA or ROCm for this path.

Reuse the expensive work

Compiled pipelines, working buffers and attention caches support repeated inference. Supported paths reuse recorded command plans to reduce repeated submission setup. The model and hardware determine which optimizations apply.

More than a chat backend

Different jobs.
Shared infrastructure.

Suite uses VulRunWG for local AI workflows. Familiar uses it for in-game assistance, including an integrated-GPU configuration that leaves the dedicated GPU rendering the game.

The Far Off Up and Rules for Thee use dedicated WGSL kernels for game effects. Particle motion and flocking are parallel compute operations, not language-model inference. Each effect has its own shader work.

The external-compute interface provides a scoped device boundary: clients own their shaders and buffers, and submit work under the runtime’s backend lock. Existing game integrations demonstrate that boundary; they do not imply unrestricted support for every GPU workload.

See the laptop workload split

Measured, with the conditions attached

Concrete work. Recorded results.

Two results from documented September 2026 development runs on an RTX 5080 with 16 GB VRAM and 32 GB system memory.

Language generation

150.4tokens / second

Qwen2.5 7B · Q4_K_M

55-token prompt, 128 greedy output tokens. Median of three after warm-up; first four token times excluded. September 28, 2026.

Image generation

9.10seconds

SDXL base · 512 × 512 · 20 steps

INT8 weights; warm render, model loading excluded. Median of process medians. September 24, 2026.

These are historical results from separate tests, not current-release benchmarks or promises for other hardware. Text decode speed excludes prompt processing and loading. Image timing excludes loading. They do not measure Familiar’s response latency.

Read the measurements and methods

Connect locally

A familiar API.
Specific supported behavior.

Suite’s local HTTP server exposes OpenAI-compatible endpoints for supported chat, embeddings, image and speech workflows. Point a compatible client at the local address and use an installed model ID.

Compatibility is not a promise that every cloud feature or client option is implemented. Check the endpoint, model, request shape and limitations for the workflow you need.

The existing Developer Tools area provides the local API address, model IDs, Python/JavaScript/C# examples and a testing console. The server and those controls belong to VulRunWG: Suite. The VulRunWG runtime provides Python interfaces and model execution; installing the runtime alone does not install Suite’s HTTP server.

Request interface documentation

The licensing boundary

Build with it. Know what you ship.

Established policy direction. Final legal agreements are still awaiting approval.

Free use

Connect to a separate runtime.

Local development, evaluation, testing and API access are free. Commercially sold software may connect to a separately installed VulRunWG runtime.

Ordinary business use and permitted internal deployment are also free. No subscription, account, activation fee or per-seat charge.

Separate agreement

Include the runtime in a sold product.

Embedding, bundling or redistributing VulRunWG as part of a commercially sold product requires a separate commercial license.

Free use does not grant unrestricted public redistribution. Model publishers and third-party components retain their own terms.

Ask about distribution

Keep the evidence attached

Numbers with context.

Hardware, model revisions, settings and limits belong beside a performance result. The historical technical pages retain their recorded results. They are development measurements, not current-release certification. Clean-machine acceptance, current model and hardware checks, and final release approvals remain to be completed.