World Model IndexEvaluate a world model
Use a repeatable task, exact version, and recorded billing conditions. The steps below are a proposed evaluation workflow.
1. Confirm access and deployment
Open the official links on the model profile. Confirm API availability, regional access, queue limits, model version, and output rights. For local models, check checkpoint availability, dependencies, license, and GPU memory.
Public code, downloadable weights, hosted inference, and commercial permission are separate facts. For example, the HY-WorldPlay repository documents variant-specific inference configurations; use the configuration for the checkpoint you plan to run.
Sources: HY-WorldPlay official repository · Matrix-Game 3.5 official repository
2. Match the task to the output
Use action scripts for navigable environments, prompt-change sequences for continuous video, matched input images for clips, and viewpoint changes plus export checks for 3D assets.
3. Measure response and continuity
Record time to first usable output, input-to-visible-response time, generated and delivered fps, stalls, and dropped frames. Report the device, network region, resolution, warmup, and trial count.
For scene recall, follow a recorded route away from and back to a reference object. Preserve the actions, seed where supported, duration, and output recording. Repeat the route before drawing conclusions.
4. Record the billing clock
Track ready session time, active streaming time, and generated output separately. Reserved GPU sessions can accrue charges while idle. Check minimums and release sessions after the test.
Sources: Reactor pricing and billing · H3 Max Director endpoint
5. Publish reproducible evidence
Keep raw results, failed runs, exact versions, dates, inputs, and links to recordings. Describe provider-reported claims separately from your observations. Use blinded, repeated comparisons for visual preference.
Download the blank test template · Read the benchmark protocol