MotionBench
The official leaderboard hasn’t scored a model released after Claude Opus 4.5 (November 2025). Labs report their own numbers for 6 newer models, such as Seed2.0 Pro at 75.2%.
Models answer questions about fine-grained motion in video clips across six motion question types. Scoring is accuracy.
Test answers are hidden and scored through the leaderboard, while the dev set has public answers.
- Made by
- Zhipu AI and Tsinghua University (THUDM)
- Released
- January 2025
- Kind
- Video understanding
- Level
- Everyday
- Scored
- Checked automatically
- Answers
- Multiple choice
- Tasks
- Public, plus a private test set
- Measures
- Capability
- Language
- English
- Interaction
- Single turn
- Refresh
- One fixed set
- Run by
- A host
- In views
- Vision
- Strongest model tested
-
Gemini 3 Pro #56 on the capability index
- Last in a launch post
- October 2026, dropped before it saturated
- Last model added
- October 2026
- Scores refreshed
- November 20th, 2026
In the authors’ words
In recent years, vision language models (VLMs) have made significant advancements in video understanding. However, a crucial capability - fine-grained motion comprehension - remains under-explored in current benchmarks. To address this gap, we propose MotionBench, a comprehensive evaluation benchmark designed to assess the fine-grained motion comprehension of video understanding models. MotionBench evaluates models' motion-level perception through six primary categories of motion-oriented question types and includes data collected from diverse sources, ensuring a broad representation of real-world video content. Experimental results reveal that existing VLMs perform poorly in understanding fine-grained motions. To enhance VLM's ability to perceive fine-grained motion within a limited sequence length of LLM, we conduct extensive experiments reviewing VLM architectures optimized for video feature compression and propose a novel and efficient Through-Encoder (TE) Fusion method. Experiments show that higher frame rate inputs and TE Fusion yield improvements in motion understanding, yet there is still substantial room for enhancement. Our benchmark aims to guide and motivate the development of more capable video understanding models, emphasizing the importance of fine-grained motion comprehension. Project page: https://motion-bench.github.io .
Seed2.0 Pro leads with 75.2%
Best score per model as labs report them in their own cards, drawn striped. Labs pick their own settings, so these compare less cleanly than independent runs.
Gemini 3 Pro leads with 70.4% across all independent runs
Best score per model across the official leaderboard, Epoch AI’s own runs and any board Epoch AI copied. Dotted bars are reruns by Epoch AI or another board, and each row names its source.
Seed2.0 Pro leads with 75.2%, claims included
Best score per model from every source, including labs’ own numbers and claims Epoch AI collected from reports and papers. Labs pick their own settings, so a claim can beat an independent run of the same model. Striped bars are claims, and dotted bars are reruns by Epoch AI or another board.
- 1
Seed2.0 Pro 75.2%
- 2
Seed2.1 Pro 74.9%
- 3
Seed2.0 Lite (0428) 72.4%
- 4
Seed1.8 70.6%
- 5 Kimi K2.5 70.4%
- 6
Seed1.5-VL 68.4%
- 7
MiniCPM-V-4.7-35B-A3B 68.1%
- 8 GLM-4.6V 63.6%
- 9 GLM-4.5V 62.4%
- 10 GLM-4.1V-9B-Thinking 59%
Source: Labs’ model cards
- 1
Gemini 3 Pro frames: - 70.4%
- 2 GPT-5.1 frames: - 66.2%
- 3 Claude Opus 4.5 frames: - 65.9%
- 4 GLM-4.5V frames: 2fps 62.4%
- 5 Claude Sonnet 4.5 frames: - 61.1%
- 6
InternVL2_5-78B frames: 16 61%
- 7
Seed1.5-VL frames: 2fps 59.4%
- 8 GLM-4.1V-9B-Thinking frames: 2fps 59.4%
- 9 GPT-5 Chat frames: 2fps 59.4%
- 10
Qwen2.5-VL-72B frames: 64 56.1%
- 11 Kimi-VL-A3B-Thinking-2506 frames: 2fps 54.3%
- 12
InternVL2-40B frames: 16 54%
- 13 MiniCPM-V-2_6 frames: 64 53%
Source: Official MotionBench leaderboard
- 75.2
Seed2.0 Pro
- 74.9
Seed2.1 Pro
- 72.4
Seed2.0 Lite (0428)
- 70.6
Seed1.8
- 70.4
Gemini 3 Pro
- 70.4 Kimi K2.5
- 68.4
Seed1.5-VL
- 68.1
MiniCPM-V-4.7-35B-A3B
- 66.2 GPT-5.1
- 65.9 Claude Opus 4.5
- 63.6 GLM-4.6V
- 62.4 GLM-4.5V
- 61.1 Claude Sonnet 4.5
- 61.0
InternVL2_5-78B
- 59.4 GLM-4.1V-9B-Thinking
- 59.4 GPT-5 Chat
- 56.1
Qwen2.5-VL-72B
- 54.3 Kimi-VL-A3B-Thinking-2506
- 54.0
InternVL2-40B
- 53.0 MiniCPM-V-2_6
- 1
Seed2.0 Pro 75.2%
- 2
Seed2.1 Pro 74.9%
- 3
Seed2.0 Lite (0428) 72.4%
- 4
Seed1.8 70.6%
- 5
Gemini 3 Pro frames: - 70.4%
- 6 Kimi K2.5 70.4%
- 7
Seed1.5-VL 68.4%
- 8
MiniCPM-V-4.7-35B-A3B 68.1%
- 9 GPT-5.1 frames: - 66.2%
- 10 Claude Opus 4.5 frames: - 65.9%
- 11 GLM-4.6V 63.6%
- 12 GLM-4.5V 62.4%
- 13 Claude Sonnet 4.5 frames: - 61.1%
- 14
InternVL2_5-78B frames: 16 61%
- 15 GLM-4.1V-9B-Thinking frames: 2fps 59.4%
- 16 GPT-5 Chat frames: 2fps 59.4%
- 17
Qwen2.5-VL-72B frames: 64 56.1%
- 18 Kimi-VL-A3B-Thinking-2506 frames: 2fps 54.3%
- 19
InternVL2-40B frames: 16 54%
- 20 MiniCPM-V-2_6 frames: 64 53%
Sources: Labs’ model cards and official MotionBench leaderboard
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 75.2% | Feb 2026 | Lab model card | |
| 2 | | 74.9% | Jun 2026 | Lab model card | |
| 3 | | 72.4% | Apr 2026 | Lab model card | |
| 4 | | 70.6% | Dec 2025 | Lab model card | |
| 5 | Kimi K2.5 | 70.4% | Jan 2026 | Lab model card | |
| 6 | | 68.4% | May 2025 | Lab model card | |
| 7 | | 68.1% | Oct 2026 | Lab model card | |
| 8 | GLM-4.6V | 63.6% | Jan 2026 | Lab model card | |
| 9 | GLM-4.5V | 62.4% | Aug 2025 | Lab model card | |
| 10 | GLM-4.1V-9B-Thinking | 59% | Jul 2025 | Lab model card |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | 70.4% | frames: - | Nov 2025 | Official leaderboard | |
| 2 | GPT-5.1 | 66.2% | frames: - | Nov 2025 | Official leaderboard |
| 3 | Claude Opus 4.5 | 65.9% | frames: - | Nov 2025 | Official leaderboard |
| 4 | GLM-4.5V | 62.4% | frames: 2fps | Aug 2025 | Official leaderboard |
| 5 | Claude Sonnet 4.5 | 61.1% | frames: - | Sep 2025 | Official leaderboard |
| 6 | 61% | frames: 16 | Dec 2024 | Official leaderboard | |
| 7 | 59.4% | frames: 2fps | May 2025 | Official leaderboard | |
| 8 | GLM-4.1V-9B-Thinking | 59.4% | frames: 2fps | Jul 2025 | Official leaderboard |
| 9 | GPT-5 Chat | 59.4% | frames: 2fps | Aug 2025 | Official leaderboard |
| 10 | 56.1% | frames: 64 | Jan 2025 | Official leaderboard | |
| 11 | Kimi-VL-A3B-Thinking-2506 | 54.3% | frames: 2fps | Jun 2025 | Official leaderboard |
| 12 | 54% | frames: 16 | Jul 2024 | Official leaderboard | |
| 13 | MiniCPM-V-2_6 | 53% | frames: 64 | Aug 2024 | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | 75.2% | Feb 2026 | Lab model card | ||
| 2 | 74.9% | Jun 2026 | Lab model card | ||
| 3 | 72.4% | Apr 2026 | Lab model card | ||
| 4 | 70.6% | Dec 2025 | Lab model card | ||
| 5 | 70.4% | frames: - | Nov 2025 | Official leaderboard | |
| 6 | Kimi K2.5 | 70.4% | Jan 2026 | Lab model card | |
| 7 | 68.4% | May 2025 | Lab model card | ||
| 8 | 68.1% | Oct 2026 | Lab model card | ||
| 9 | GPT-5.1 | 66.2% | frames: - | Nov 2025 | Official leaderboard |
| 10 | Claude Opus 4.5 | 65.9% | frames: - | Nov 2025 | Official leaderboard |
| 11 | GLM-4.6V | 63.6% | Jan 2026 | Lab model card | |
| 12 | GLM-4.5V | 62.4% | Aug 2025 | Lab model card | |
| 13 | Claude Sonnet 4.5 | 61.1% | frames: - | Sep 2025 | Official leaderboard |
| 14 | 61% | frames: 16 | Dec 2024 | Official leaderboard | |
| 15 | GLM-4.1V-9B-Thinking | 59.4% | frames: 2fps | Jul 2025 | Official leaderboard |
| 16 | GPT-5 Chat | 59.4% | frames: 2fps | Aug 2025 | Official leaderboard |
| 17 | 56.1% | frames: 64 | Jan 2025 | Official leaderboard | |
| 18 | Kimi-VL-A3B-Thinking-2506 | 54.3% | frames: 2fps | Jun 2025 | Official leaderboard |
| 19 | 54% | frames: 16 | Jul 2024 | Official leaderboard | |
| 20 | MiniCPM-V-2_6 | 53% | frames: 64 | Aug 2024 | Official leaderboard |
The benchmark's own leaderboard has 24 entries, 13 of them models we track. Switch the scores above to rank them together.
Rankings by part
We read 6 parts from the benchmark’s own leaderboard, each ranking models on one part of the test, such as Motion recognition, Location-related motion and Camera motion. Gemini 3 Pro leads 5 of them, and Claude Opus 4.5 leads the other one. Arrows show how far a model moved from its overall rank among the models we track.
Motion recognition
13 models we track have a score here.
- 1
Gemini 3 Pro 70.9%
- 2 GPT-5.1 69.6%
- 3 Claude Opus 4.5 66.5%
- 4
InternVL2_5-78B ↑2 65.2%
- 5 GLM-4.5V ↓1 65.1%
- 6 Claude Sonnet 4.5 ↓1 63.2%
- 7 GLM-4.1V-9B-Thinking ↑1 62.5%
- 8
Seed1.5-VL ↓1 61.4%
- 9 GPT-5 Chat 61.4%
- 10
Qwen2.5-VL-72B 59.5%
Location-related motion
13 models we track have a score here.
- 1
Gemini 3 Pro 73.9%
- 2 Claude Opus 4.5 ↑1 71%
- 3 GPT-5.1 ↓1 68.8%
- 4
Seed1.5-VL ↑3 66.7%
- 5 GPT-5 Chat ↑4 66.7%
- 6 Claude Sonnet 4.5 ↓1 66.5%
- 7 GLM-4.5V ↓3 65.2%
- 8
InternVL2_5-78B ↓2 64.6%
- 9 GLM-4.1V-9B-Thinking ↓1 63%
- 10
Qwen2.5-VL-72B 59%
Camera motion
13 models we track have a score here.
- 1 Claude Opus 4.5 ↑2 71%
- 2
Gemini 3 Pro ↓1 66.7%
- 3 Claude Sonnet 4.5 ↑2 59.2%
- 4 GLM-4.5V 58.7%
- 5
Seed1.5-VL ↑2 58.6%
- 6 GPT-5 Chat ↑3 58.6%
- 7
InternVL2_5-78B ↓1 55.6%
- 8 GLM-4.1V-9B-Thinking 53.9%
- 9 GPT-5.1 ↓7 53.6%
- 10
InternVL2-40B ↑2 49%
Motion-related objects
13 models we track have a score here.
- 1
Gemini 3 Pro 83.4%
- 2 Claude Opus 4.5 ↑1 79.7%
- 3 GLM-4.5V ↑1 77.4%
- 4 GPT-5.1 ↓2 76.6%
- 5 GLM-4.1V-9B-Thinking ↑3 76.2%
- 6
InternVL2-40B ↑6 76%
- 7
InternVL2_5-78B ↓1 75.7%
- 8 Claude Sonnet 4.5 ↓3 74.3%
- 9
Seed1.5-VL ↓2 73.8%
- 10 GPT-5 Chat ↓1 73.8%
Action order
13 models we track have a score here.
- 1
Gemini 3 Pro 55%
- 2 Claude Opus 4.5 ↑1 55%
- 3 GPT-5.1 ↓1 54.6%
- 4 Claude Sonnet 4.5 ↑1 54.4%
- 5
Seed1.5-VL ↑2 49.4%
- 6 GPT-5 Chat ↑3 49.4%
- 7 GLM-4.5V ↓3 48.3%
- 8
InternVL2_5-78B ↓2 46.9%
- 9 GLM-4.1V-9B-Thinking ↓1 46.2%
- 10
Qwen2.5-VL-72B 45.5%
Repetition count
13 models we track have a score here.
- 1
Gemini 3 Pro 61.2%
- 2 GPT-5.1 56.7%
- 3 GLM-4.5V ↑1 40.4%
- 4 Claude Opus 4.5 ↓1 36.9%
- 5
InternVL2_5-78B ↑1 33.7%
- 6 GLM-4.1V-9B-Thinking ↑2 33.5%
- 7 MiniCPM-V-2_6 ↑6 33%
- 8
InternVL2-40B ↑4 30%
- 9 Claude Sonnet 4.5 ↓4 29.4%
- 10
Qwen2.5-VL-72B 28.7%
3 models have held number one in labs’ own numbers since its release in January 2025 4 models have held number one across all independent runs since its release in January 2025 5 models have held number one with claims included since its release in January 2025
Lab claims count from the day the card came out, independent runs from the model’s release. We draw one row each for the benchmark’s own leaderboard and what labs claim in their cards.
Sources: Official MotionBench leaderboard and labs’ model cards
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Sources: Official MotionBench leaderboard and labs’ model cards
Labs headline it less often now, last in October 2026
Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out. ByteDance Seed accounts for 5 of the 10 releases that named it since 2024.
| Date | Launch post | Score | As the post writes it |
|---|---|---|---|
| Oct 7, 2026 | 68.1 | MotionBench | |
| Apr 28, 2026 | 72.4 | MotionBench | |
| Feb 14, 2026 | 75.2 | MotionBench | |
| Jan 27, 2026 | Kimi K2.5 | 70.4 | MotionBench |
| Dec 18, 2025 | 70.6 | MotionBench | |
| Aug 11, 2025 | GLM-4.5V | 62.4 | MotionBench |
| Jul 1, 2025 | GLM-4.1V-9B-Thinking | 59 | MotionBench |
| May 12, 2025 | 68.4 | MotionBench |
8 system cards report it, from 3 labs
Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| Jun 23, 2026 | | 74.9 | MotionBench (Table 10) | Capabilities |
| Mar 21, 2026 | | 70.6 | MotionBench (Table 3) | Capabilities |
| Feb 14, 2026 | | 75.2 | MotionBench (Table 9) | Capabilities |
| Jan 30, 2026 | Kimi K2.5 technical report | 70.4 | MotionBench (Table 4) | Capabilities |
| Jan 1, 2026 | GLM-4.6V technical report | 63.6 | MotionBench, GLM-4.6V thinking (Table 2) | Capabilities |
| Aug 13, 2025 | GLM-4.5V technical report | 62.4 | MotionBench, GLM-4.5V thinking (Table 2) | Capabilities |
| Jul 1, 2025 | GLM-4.1V-9B-Thinking technical report | 59 | MotionBench, GLM-4.1V-9B-Thinking thinking (Table 2) | Capabilities |
| May 11, 2025 | | 68.4 | MotionBench, thinking (Table 7). Non-thinking 68.4 | Capabilities |
The lab's number comes out ahead in 1 of 2 pairs that differ
Each score a lab reported for its own model, beside the same model's best score on the leaderboard. Gaps mostly come from different harnesses, effort levels and task subsets. 1 of them match, which usually means one side copied the other. Every benchmark.
| Model | Lab says | Board says | Gap |
|---|---|---|---|
| | 68.4% MotionBench | 59.4% Leaderboard, frames: 2fps | +9.0 pts |
| GLM-4.1V-9B-Thinking | 59% MotionBench | 59.4% Leaderboard, frames: 2fps | −0.4 pts |
| GLM-4.5V | 62.4% MotionBench | 62.4% Leaderboard, frames: 2fps | Same |
More in Audio and video
- Video-MMMU In use 300 expert-level lecture videos with questions testing perception, comprehension and adaptation of knowledge. Accuracy.
Qwen 3.8 Max 88.7%
- LVBench In use Hours-long video understanding with multiple-choice questions on reasoning, grounding and summarization.
Gemini 4 Argon 91.7%
- MMVU In use 3,000 expert-annotated questions on videos across 27 disciplines. Multiple-choice and open-ended accuracy.
Qwen 3.8 Max 82.4%
- MLVU In use Multi-task long video understanding (minutes to hours) with multiple-choice and generation tasks. Accuracy.
Qwen 3.8 Max 90.8%
- FLEURS In use Models transcribe read speech in 102 languages built from FLoRes-101 sentences. Model cards usually report word error rate, so lower scores are better.
Phi-4-multimodal 4.00%
- VoiceBench In use Spoken instructions with speaker, environment and content variation. Voice assistant responses scored by GPT judge or rules.
Qwen3.5-Omni-Plus 93.1%
Data audit
All 11 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access Public tasks, plus a private test set.
- Pass Confidence Checked against a primary source.
Scores 4 pass 1 note
- Pass Independent runs Scores from the official leaderboard (24 entries), launch posts (8 scores) and system cards (8 scores).
- Pass Models scored 20 models we track, plus 11 other entries on its leaderboard.
- Pass Newest score The newest is from October 2026, for MiniCPM-V-4.7-35B-A3B.
- Note Frontier models None of the top 10 has a score. The strongest tested is Gemini 3 Pro (#56).
- Pass Headroom The best is 75.2%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on motion-bench.github.io: 13 rows for models we track and 11 other entries at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 2 pass
- Pass Lab claims All 3 pairs agree within 10 points.
- Pass Known issues No published audit found problems.