@readmodelmag
MODEL

Tools

Vork: a headless 3D/VFX engine built for AI agents

SummaryNov 6, 2026: Ravi Parikh / Voyager ships Vork research preview — agents write scenes as code (Rust + Slang; Vulkan/Metal). Getting-started help ~2,500 tokens. Vendor demo: 1M self-steering birds simulated+rendered in 97s on one NVIDIA L4 (after setup/warmup).

· MODEL Editors · 2 min read

Some links may be affiliate links. Disclosure

MODEL cartoon on a cream background. Caption: Humans get a thousand settings. Agents get a 2,500-token engine. Left panel: HUMAN TOOLS — cascading generic Settings dropdowns (Render Engine, Samples, Light Paths, etc.), stressed stick figure drowning in menus. Right panel: AGENT + VORK — code block labeled ~2,500 tokens, cute robot at a keyboard, flock of bird silhouettes, badge 1,000,000 birds · 97s · 1× L4. Watermark MODEL · readmodelmag.com.
Humans drown in DCC settings; agents get a small scene-as-code engine. Facts from voyager.so/vork + creator X; vendor claims; not tested by MODEL. Illustration by MODEL. Original. Image source.

Vork is a headless 3D and VFX engine built for agents, announced Nov 6, 2026 by Ravi Parikh (@ravisparikh (opens in a new tab)) / Voyager. Per voyager.so/vork (opens in a new tab): an agent writes a scene as code, renders it, inspects the result, and revises. Stack: Rust + Slang, GPU via Vulkan / Metal. It is a research preview — source-available and free during the preview.

Small surface for agents. The site says getting-started help is about 2,500 tokens, including a complete first scene. Ordinary shots need scene data; programming hooks cover unusual cases. Blender / Houdini / Unreal / Nuke are framed as human tools with thousands of settings; Vork’s kernel holds what agents can’t write on the fly (geometry, ray queries), and the rest is code the agent writes as it goes.

Million-bird demo (vendor numbers). Site: a million birds, each steering itself, simulated and rendered in 97 seconds on one GPU. Timing nuance on the same page: the eight-second skim took 97s after 244s setup + 166s scene warmup on an NVIDIA L4 (asset prep separate). EEVEE is projected 13–18 minutes draw-only at the same samples (~8–11×), flock handed to Blender — quality/material/motion differ. Camera + nine foreground birds authored; the million simulated birds don’t flap. Astra wrote the shot; flocking sim by GPT-6.1 Sol earlier. Creator X also claimed 2× more token-efficient vs Blender/Unreal/Houdini and up to 24× faster rendering — those are creator claims, not MODEL measurements.

Not a video model. Moving parts are rendered 3D, or real footage + rendered 3D (tracking/cutouts via tools like OpenCV / SAM; one car-shot backdrop used an image-model painting on a 3D card).

Install (per site). Windows/Linux with a recent NVIDIA GPU; Apple Silicon with Metal (bundled Slang targets macOS 26; older macOS unqualified).

Research-preview limits (must-know). Fog/volumes not in yet; heavy interiors can abort on Mac; path tracer slower than Cycles in measured cases; Final approximates indirect lighting (more samples cannot restore omitted transport). Run vork help limits before committing to a shot.

Who it's for: founders building agent loops that need controllable 3D/VFX output, and teams tired of driving DCC UIs from an LLM.

Caveats. All performance and token figures above are vendor / creator-reported. MODEL has not installed or benchmarked Vork.

Disclosure: Not sponsored. No affiliate links. Not tested by MODEL.

MODEL Editors

Sources

  1. Vork by Voyager

    Voyager / Ravi Parikh

  2. Ravi Parikh announces Vork

    X / @ravisparikh

  3. voyager-so/vork

    GitHub

FREE · EVERY WEEK

The top AI tools this week, tested.

Every week MODEL's editors pick the AI tools worth your time, try them hands-on, and tell you what's good, what's not, and who it's for. Free, in your inbox.

Free. Unsubscribe anytime. Our picks are editorial and can't be bought.