Seedance 2.5 tops an AI video physics test at 57.76 out of 100

ByteDance's Seedance 2.5 finished first in a new physics benchmark for AI video, with an overall score of 57.76 out of 100. The benchmark, World Models' Last Exam in Physics, was submitted to arXiv on October 6, 2026 by Mingju Gao and ten co-authors. It gives eight image-to-video models 40 controlled tasks and four tries each, 1,280 clips in all, then measures whether the physics in each clip holds up.

The method is the point. Instead of asking a vision model whether a video looks right, the authors track objects, fit lines and read liquid levels, then compare what they measure to the relationship the scene should satisfy. A pendulum's period should not change when you swap the bob's mass. A reflected ray should leave at the angle it arrived. Each clip gets a consistency score, and only clips scoring 80 or higher on consistency count their physical score toward the composite, which weighs physics at 0.85 and consistency at 0.15. Clearing that screen is not a pass on physics: every free-fall clip passed it, and the eight models still averaged 26.61 of 100 on that task.

For a creator, physical consistency is the set of things a viewer checks without thinking. Does the ball bounce to a lower height each time, does poured water settle at a level surface, does a dropped glass break into pieces that stay where they land. When a clip gets those wrong, the shot reads as fake, and a product demo or a bounce that anchors an ad stops being usable in a final cut. This ranking covers the eight models the paper scored, ordered by its measured composite scores. No hands-on tests were run for it. Seedance 2.5 is the only closed model in the set; the other seven are open weights or derivatives of open weights, with MiniMax saying it will release the H3 weights. The set leaves out Google's Veo, Kuaishou's Kling, OpenAI's Sora, Runway and LTX, so it is not a ranking of the whole field.

8. CogVideoX1.5-5B (Zhipu AI and THUDM)

CogVideoX1.5-5B is the oldest and smallest model in the set, an open-source 5B model that Zhipu AI and Tsinghua's THUDM group released in November 2024, per its Hugging Face model card. It finished last at 18.81 of 100, 13.16 points below HunyuanVideo-1.5 in seventh. That gap is the widest between any two neighbouring models on the list, which makes the two generations of models in this test read as two different eras of the same problem.

Its best use is cheap local generation and a fixed baseline to measure newer models against. The catch is that its low score is partly a scoring artifact. Under the benchmark's gating, a video that fails the consistency screen counts its physical term as zero, so a model whose clips are hard to measure loses points it never gets a chance to earn. It was tested at 832x480 and 81 frames.

7. HunyuanVideo-1.5 (Tencent)

Tencent's HunyuanVideo-1.5 is an 8.3B open-weight model built to run on consumer GPUs, with text-to-video and image-to-video in one package. It finished seventh at 31.97 of 100, 0.28 behind LingBot-Video in sixth, a near-tie.

The finding worth keeping is how it placed. HunyuanVideo-1.5 was tested at 1264x720 and 129 frames, the longest sequence in the set, and it landed near the bottom on physics. Visual quality and physical consistency are scored separately here, and a model can win the first while losing the second. Its strength is efficient local generation; its weakness is that producing a clean-looking shot says nothing about whether the motion inside it obeys a physical relationship.

6. LingBot-Video (Robbyant, Ant Group)

LingBot-Video comes from Robbyant, the Ant Group embodied-AI company, and its maker bills it as the first open-source Mixture-of-Experts video generation model dedicated to embodied intelligence. It runs a 30B-A3B design, 30 billion total parameters with about 3 billion active per token, trained on 70,000-plus hours of embodiment-oriented data, per the LingBot-Video paper. It finished sixth at 32.25 of 100.

The story here is the gap between the pitch and the measurement. LingBot-Video is trained with a reward system that rewards physical rationality, and it sits in the bottom half of this benchmark. Physical grounding as a training goal and physical consistency as a measured outcome are not the same thing yet. It edges HunyuanVideo-1.5 by 0.28, so the two are effectively level.

5. Wan2.2-I2V-A14B (Alibaba)

Alibaba's Wan2.2-I2V-A14B is a Mixture-of-Experts image-to-video model with 27 billion total parameters and 14 billion active per step, released under Apache 2.0 and supporting 480P and 720P. It finished fifth at 35.66 of 100. It is the base model that VBVR-Wan2.2, one rank above it, was fine-tuned from.

That relationship is the useful detail. The two models share an architecture, and the fine-tune scores 37.79 against the base model's 35.66, a 2.13-point gain from post-training alone. Physics scores can be moved after the fact, so a bigger model is not the only route to better physics. Its best use is license-free local generation, and its catch is that the model next to it on this list does the same job slightly better for free.

4. VBVR-Wan2.2 (Video-Reason team)

VBVR-Wan2.2 is a fine-tune of Alibaba's Wan 2.2 from the Video-Reason team behind the Very Big Video Reasoning suite, and it is open source, per the VBVR leaderboard and the VBVR paper. It finished fourth at 37.79 of 100, a clear step up from the base Wan2.2-I2V-A14B below it.

Its standout result is in optics. On the light-reflection task (P18), VBVR-Wan2.2 scored 95.75, and Seedance 2.5 scored 96.15, the only two models in the run to clear 90. Every other model scored at most 49.82, and the paper notes that reflected rays are visible in the VBVR and Seedance outputs and missing in several others. A model that gets mirrors and light right is worth knowing about if your shot has glass, water or chrome in it. Its catch is that the fine-tune helps one physical domain far more than the others.

3. Cosmos3-Super (NVIDIA)

NVIDIA's Cosmos3-Super is the 64B member of its Cosmos 3 family of omnimodal world models, released openly on Hugging Face and built for physical AI, robotics and simulation rather than for creative video. It finished third at 42.46 of 100, the best score of any open-weight model and 4.67 points above VBVR-Wan2.2.

Its placement is the headline for the open-weights camp, and it is also a caution. Cosmos3-Super is the model here most focused on reasoning about the physical world, and it trails MiniMax H3 by 12.43 and Seedance 2.5 by 15.30. Open weights buy you control and a license you can build on; in this run they did not buy the top physics score. Its best use is simulation and robot-adjacent work where an open, inspectable world model matters more than a few points of video fidelity.

2. MiniMax H3 (MiniMax)

MiniMax H3 is a general-purpose multimodal model that renders 2K video up to 15 seconds with native stereo sound, and MiniMax says it will open the model weights. It finished second at 54.89 of 100, 2.87 behind Seedance 2.5. It led the benchmark's easy task group and matched Seedance at the top of the charged-sphere equilibrium task (P31), where it scored 97.24 against Seedance's 98.34 while every other model scored at most 15.00.

Its weak spot is the hard end. On tasks that chain events together, collisions, motion reversal and melting, MiniMax H3 averaged 19.87, less than Seedance 2.5's 37.76. MiniMax's own launch post says H3 delivers 2K at less than a third of the per-second price of mainstream models and 768p at less than half of their 720p, which makes it the value pick of the two leaders for simple, stable shots. The catch is that the shots most likely to expose bad physics are the ones it handles least well.

1. Seedance 2.5 (ByteDance)

Seedance 2.5 is ByteDance's newest video model and the only closed model in the set. It tops the benchmark at 57.76 of 100. Its edge shows up most in the hard cases: it took the highest task-averaged hard score at 37.76, roughly double MiniMax H3's 19.87, and it posted the top result on charged-sphere equilibrium (98.34) and on light reflection (96.15).

What it is good at, per ByteDance's launch post, is long single-pass generation: up to 30 seconds of audio-video in one pass with native audio, extendable across rounds, and up to 50 multimodal references. ByteDance also names the limit this benchmark measures. The post says the model still has "room for improvement, particularly regarding the physical plausibility of complex motions." A top score of 57.76 means roughly four in ten measured physical relationships still fail, so even the best model here produces shots a creator should plan to re-roll.