MiniMax H3 scores 54.89 out of 100 in the published results of World Models’ Last Exam in Physics, behind Seedance 2.5 at 57.76 and ahead of the other six evaluated models. Those numbers come from the researchers’ public results table, not a benchmark conducted by H3Vid.
The paper was submitted on October 6, 2026, and MiniMax highlighted it on October 8. The useful question is what the score means for a shot you actually want to make. A strong position within this comparison does not certify that every generated scene obeys physics. The cover is a generated concept illustration, not a benchmark frame or scientific measurement.
What the exam measures
The benchmark combines 40 controlled tasks across physical domains, using an initial image, a prompt, and predefined measurable criteria. Its experiments cover eight models and 1,280 videos. The research project provides context and examples.
This differs from asking viewers which clip looks more cinematic. The goal is to examine observable relationships: motion, geometry, or changes in a physical scene. A polished texture can coexist with incorrect timing, while an unattractive render can still depict the intended relationship.
For creators, treat the exam as a diagnostic resource. Find the task closest to your planned shot before interpreting an overall score. A reflection scene and a falling-object scene can expose different weaknesses even when they use the same model.
Why a check mark is not a physics pass
The repository reports a composite score with a consistency and observability component, plus a separately measured physical component. Physical contribution is gated by an observability threshold. Its table’s check mark indicates that every available video for that model–task pair meets the specified consistency threshold; it does not mean perfect physical behavior.
Free fall makes this distinction concrete. The table lists MiniMax H3 at 25.84 with a check mark, while its bounce-height-decay task is 88.19 with a check mark. Both sets are observable enough under the screen, but their physical scores differ substantially.
Do not translate a check mark into “this action always works.” Instead, ask two questions: is the action visible and trackable, and does its measured behavior satisfy the task’s criteria? A clip can answer the first question positively while failing the second.
Avoid turning a small overall gap into a universal ranking
The aggregate difference between 57.76 and 54.89 is 2.87 points on this reported scale. That subtraction is not a probability that one model wins your next shot, and it is not a quality difference measured in percent.
The comparison applies to the tested configurations and tasks. It does not automatically include H3 Max, H3 Max Turbo, community LoRAs, or every cloud provider’s preprocessing. Keep model identity and generation settings attached to any reused result.
If choosing a production workflow, add the criteria the benchmark does not settle for you: useful duration, reference control, sound, acceptable retries, and delivery format. Use the physics result to identify shots needing closer review rather than replacing the whole selection process with one number.
Review a falling or bouncing object
For your own shot, begin with a fixed camera and one clearly visible object. This is a proposed review method, not a reproduction of the benchmark evaluator. Keep the untouched generated clip and record its frame rate.
Scrub from release to impact. Note whether the object accelerates plausibly, changes shape, pauses unexpectedly, or disappears behind an ambiguous blur. With a fixed view, successive positions are easier to compare than when the camera also accelerates.
For a bounce, inspect the instant of contact as well as the arc. Does the object penetrate the floor, change its scale, or gain energy unexpectedly? Follow several rebounds if the shot includes them. A convincing first impact can conceal an implausible continuation.
These checks are editorial inspection. Quantitative claims require calibrated measurements and the task’s assumptions; perspective and camera motion can make screen-space distances misleading. Do not attach a homemade numeric “physics score” to casual scrubbing.
Review reflections, fluids, and continuity separately
In a reflected scene, compare the subject and its reflection across time. Check their relative motion, occlusion, and contact with the reflecting surface. A reflection that looks attractive in one frame can drift away from its source in the next.
For a liquid shot, decide what should remain consistent before generation: the container’s geometry, the liquid’s approximate amount, and the intended movement. Watch for sudden level changes or an object that behaves differently after an occlusion. These are useful rejection signals, not proof that an accepted clip solves fluid dynamics.
Keep physical review distinct from identity and temporal continuity checks. A stable character can move implausibly; a physically believable object can change appearance. Writing separate acceptance criteria makes it easier to explain a failed take and choose a targeted retry.
How to use the result without overclaiming
Use the public task descriptions and inputs when planning a genuine benchmark reproduction. The repository documents generation and evaluation commands, environments, and custom-model configuration. We have not executed them or repeated the reported rankings.
For ordinary filmmaking, start with a short, inspectable shot and change one variable when retrying. If you intend to slow footage or cut before an impact, describe that editorial choice explicitly; it does not demonstrate that the original motion was physically accurate.
The next action is to identify the physical event in your storyboard, select a relevant benchmark task, and define what you will inspect in the finished clip. The H3 motion-control page can help structure an action and camera prompt, while the H3 Max page explains a separate product variant. Neither should be presented as having inherited the base H3 benchmark score without a matching test.