Models

Use This Page

Use this page when you need to pick a model, check whether it can run on your machine, or understand what a support label means. The table names below are catalog groups, not separate commands. For day-to-day use, start from the CLI:

Run inference

With a prepared requests.jsonl, generate artifacts via the pipeline runner:

worldfoundry-eval evaluate \
  --mode model \
  --model-id <model-id> \
  --model-runner worldfoundry:pipeline \
  --model-manifest-dir worldfoundry/data/models/catalog \
  --requests-path tmp/requests.jsonl \
  --output-dir tmp/model_eval \
  --metric artifact_count \
  --required-artifact video \
  --json

For the full inference flow (environment, checkpoints, Studio), see the Inference guide.

Output files

When a run fails

1. zoo model-show --include-manifest — integration status and blockers
2. zoo model-download --check-local — checkpoint, credential, and asset readiness
3. Cross-check the benchmark page for required datasets, API keys, and generated-output layout
4. Cross-check the model registry below for runner evidence on that model id

Choosing a Model

Before spending GPU time, answer these questions in order:

Reading Catalog Rows

Current coverage summary

WorldFoundry tracks model support at the manifest, runner, and benchmark-evidence levels. Use this summary to find the right model group first, then inspect the exact model id with zoo model-show.

Model source registry

This registry combines worldfoundry/data/models/catalog manifests with official paper, project-page, and source links confirmed for this docs surface. When a separate paper, project page, or code repository is not published, the closest official model card, source page, or project URL is used.

Open video models

Upstream video-generation families and in-tree runtime or metadata records. Support levels vary: some ids have verified local generation paths, others are checkpoint-gated wrappers, hosted/API-shaped entries, or provenance-only. Use catalog readiness and zoo model-show before treating a row as runnable.

World models

Interactive, camera-controlled, navigation, driving, and long-horizon simulation systems, including provenance or metadata-only rows. Capability tags are contract hints: action-controlled models should preserve action traces, camera-controlled models should expose camera metadata, and long-horizon systems need artifact naming that preserves sequence order.

3D and 4D models

3D/4D generation, reconstruction, point-cloud, geometry, and perception-prior models, including provenance or metadata-only rows. Prefer entries with explicit geometry, camera, or asset metadata when you need measurable 3D outputs; some rows still need a port plan before they are runnable in WorldFoundry.

VLA, VA, and WAM models

These models are evaluated through behavior, not image quality alone. A usable run needs observations, language or task instructions, action outputs, simulator or dataset context, and a success signal that can be mapped into normalized metrics.

openvla is the checkpoint-verified embodied-action baseline in this group. MolmoAct2 also has an in-tree WorldFoundry pipeline/runtime surface for DROID and bimanual YAM action-trace contracts, with checkpoint-backed GPU parity still pending. The former generic LeRobot policy-zoo surface is retired; policies from lerobot/src/lerobot/policies are now tracked as concrete algorithms with their own upstream sources, such as ACT, Diffusion Policy, TD-MPC, VQ-BeT, pi0/pi0.5/pi0-FAST, SmolVLA, GR00T, X-VLA, WALL-OSS, EO-1, MolmoAct2, Multi-task DiT, Gaussian Actor, and RTC. Most VLA/VA/WAM rows are currently runtime_ported: they have a WorldFoundry route contract, but task support is still bounded by checkpoint, simulator, and real inference evidence.