Omni world model for camera-controlled generation, spatial reconstruction and simulation. The research release and the access-dependent Marble 2 beta API expose different documented limits.
Each new multimodal sequence element is generated autoregressively through diffusion; input views are grounded in a shared spatial context.
Languages
Not collected
Context caching
Not collected
No published prompt-cache contract or discounted input-token tariff. Internal spatial/KV memory, reusable scene prompts and stored exports are different mechanisms.
The guide describes current distilled models using eight steps, but stable individual checkpoint identities and a model selector are not exposed by this public request schema.
No stable API model selector verified; task names and descriptive research labels are not checkpoint selectors.
Access
Account-enabled preview tasks at https://api.atlas-beta.worldlabs.ai/api/v2/. API keys belong to a project. Each enabled task has its own inputs and limits; public docs do not guarantee access for every account.
Regions
No model-specific selectable processing region or data-residency commitment verified.
Account defaults apply if omitted. An operation cannot disable account-level purge. Library uploads are not purged by this operation flag. This is not zero retention and does not replace content-use terms.
Documented for the image-to-RGBD task; no universal upload-size cap is inferred.
Endpoint pricing has not been collected. Check the provider for applicable rates.
Data & enterprise terms
Terms permit use of user content for development, benchmarking and model training. Paid users can opt out prospectively; this does not remove data from models already trained. Free and paid output rights differ. Retention is purpose-based with legal exceptions; no general zero-retention promise.
Caching can reuse earlier input processing. Availability, lifetime and billing depend on the endpoint. A cache hit still generates a new response.
01 / INITIAL INPUTProcess & retain
The system processes eligible input and retains reusable state.
02 / LATER REQUESTCheck available state
Reuse depends on matching input, session state and the cache policy.
03 / CACHE HITReuse & generate
Reusing state can reduce repeated work. Any charges follow the billing rules below.
Marble 2 beta Developer API
Not verified
How it works
Not collected
Minimum reusable input
Not collected
Lifetime & refresh
Not collected
How to enable it
Not collected
Usage fields
Not collected
Billing rules
Not collected
Limits & data boundaries
Not collected
No verified cache rate for this endpoint. An empty rate is not a zero charge.
Caching reviewed 14 Oct 2026
What the measurements say
Camera-path preference vs MiniMax H3
World Labs · vendor-reported
75% choosing Atlas
Provider evaluation using third-party human raters. Single input image with one to three cinematic motions; Atlas receives native camera inputs, competitors receive text descriptions. Sample size, confidence intervals and exact API checkpoints are not disclosed. Measures camera following, not overall model quality.
Provider evaluation using third-party human raters. Single input image with one to three cinematic motions; Atlas receives native camera inputs, competitors receive text descriptions. Sample size, confidence intervals and exact API checkpoints are not disclosed. Measures camera following, not overall model quality.
Provider evaluation using third-party human raters. Single input image with one to three cinematic motions; Atlas receives native camera inputs, competitors receive text descriptions. Sample size, confidence intervals and exact API checkpoints are not disclosed. Measures camera following, not overall model quality.
Provider evaluation using third-party human raters. Single input image with one to three cinematic motions; Atlas receives native camera inputs, competitors receive text descriptions. Sample size, confidence intervals and exact API checkpoints are not disclosed. Measures camera following, not overall model quality.
Provider evaluation using third-party human raters. Single input image with one to three cinematic motions; Atlas receives native camera inputs, competitors receive text descriptions. Sample size, confidence intervals and exact API checkpoints are not disclosed. Measures camera following, not overall model quality.
Mean absolute-relative pointmap error; lower is better. Sparse posed input images; provider reports rerunning all comparison baselines under a common protocol. Exact splits, image counts and per-example settings are not supplied. The Average row is an aggregate, not another dataset.
Mean absolute-relative pointmap error; lower is better. Sparse posed input images; provider reports rerunning all comparison baselines under a common protocol. Exact splits, image counts and per-example settings are not supplied. The Average row is an aggregate, not another dataset.
Mean absolute-relative pointmap error; lower is better. Sparse posed input images; provider reports rerunning all comparison baselines under a common protocol. Exact splits, image counts and per-example settings are not supplied. The Average row is an aggregate, not another dataset.
Mean absolute-relative pointmap error; lower is better. Sparse posed input images; provider reports rerunning all comparison baselines under a common protocol. Exact splits, image counts and per-example settings are not supplied. The Average row is an aggregate, not another dataset.
Mean absolute-relative pointmap error; lower is better. Sparse posed input images; provider reports rerunning all comparison baselines under a common protocol. Exact splits, image counts and per-example settings are not supplied. The Average row is an aggregate, not another dataset.
Mean absolute-relative pointmap error; lower is better. Sparse posed input images; provider reports rerunning all comparison baselines under a common protocol. Exact splits, image counts and per-example settings are not supplied. The Average row is an aggregate, not another dataset.
Mean absolute-relative pointmap error; lower is better. Sparse posed input images; provider reports rerunning all comparison baselines under a common protocol. Exact splits, image counts and per-example settings are not supplied. The Average row is an aggregate, not another dataset.
Mean absolute-relative pointmap error; lower is better. Sparse posed input images; provider reports rerunning all comparison baselines under a common protocol. Exact splits, image counts and per-example settings are not supplied. The Average row is an aggregate, not another dataset.
Latency and throughput depend on the host, request size, reasoning settings and load. There is no measured speed or global usage figure in this record.
The record behind the model
Why it is included
World Labs identifies its AI models as proprietary, retains their intellectual property and restricts redistribution and reverse engineering. Public APIs, generated 3D assets and the open Spark renderer do not provide the underlying model weights. Historical research-checkpoint distribution remains unverified.
