SAVU-Bench
Models answer spatial questions about real scenes using sound and video. The authors report accuracy across seven tasks.
- Made by
- Yu Chen and colleagues
- Released
- October 2026
- Kind
- Audio understanding and voice
- Level
- Everyday
- Scored
- Checked automatically
- Answers
- Multiple choice
- Measures
- Capability
- Language
- English
- Interaction
- Single turn
- Items from
- Real work and use
- Refresh
- One fixed set
- Run by
- A host
- In views
- Vision, Audio and speech
- Strongest model tested
-
Gemini 3.6 Flash #45 on the capability index
- Last model added
- July 2026
- Scores refreshed
- November 20th, 2026
Gemini 3.6 Flash leads with 55.3% among models released by July 2026 Gemini 3.6 Flash leads with 55.3% among models released by July 2026
Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. The paper reports no overall score, so this is the plain mean of its seven task accuracies. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.
The newest model with a score here came out in July 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 55.3
Gemini 3.6 Flash
- 52.7 Gemini-2.5-Pro (snapshot unspecified)
- 52.6
Qwen3.5-Omni-Plus
- 51.7
Doubao Seed 2.0 Lite
- 51.2
Qwen3.5-Omni-Flash
- 51.1 DOTS-3-Note-Preview
- 50.1
MiniCPM-o 4.5
- 49.7 Spatial-Omni SO-7B
- 49.4
Nemotron 3 Nano Omni 30B-A3B
- 48.8 video-SALMONN2+ 7B
- 48.8
Qwen2.5-Omni-7B
- 47.1 OmniVideo-7B (VITA1.5)
- 1
Gemini 3.6 Flash 55.3%
- 2 Gemini-2.5-Pro (snapshot unspecified) 52.7%
- 3
Qwen3.5-Omni-Plus 52.6%
- 4
Doubao Seed 2.0 Lite Apr 2026 51.7%
- 5
Qwen3.5-Omni-Flash 51.2%
- 6 DOTS-3-Note-Preview 51.1%
- 7
MiniCPM-o 4.5 50.1%
- 8 Spatial-Omni SO-7B 49.7%
- 9
Nemotron 3 Nano Omni 30B-A3B 49.4%
- 10 video-SALMONN2+ 7B 48.8%
- 11
Qwen2.5-Omni-7B 48.8%
- 12 OmniVideo-7B (VITA1.5) 47.1%
Source: The SAVU-Bench paper
- 1
Gemini 3.6 Flash 55.3%
- 2
Qwen3.5-Omni-Plus 52.6%
- 3
Doubao Seed 2.0 Lite Apr 2026 51.7%
- 4
Qwen3.5-Omni-Flash 51.2%
- 5
MiniCPM-o 4.5 50.1%
- 6
Nemotron 3 Nano Omni 30B-A3B 49.4%
- 7
Qwen2.5-Omni-7B 48.8%
Source: The SAVU-Bench paper
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 55.3% | Jul 2026 | Official leaderboard | |
| 2 | Gemini-2.5-Pro (snapshot unspecified) | 52.7% | Official leaderboard | ||
| 3 | | 52.6% | Mar 2026 | Official leaderboard | |
| 4 | | 51.7% | Apr 2026 | Official leaderboard | |
| 5 | | 51.2% | Mar 2026 | Official leaderboard | |
| 6 | DOTS-3-Note-Preview | 51.1% | Official leaderboard | ||
| 7 | | 50.1% | Feb 2026 | Official leaderboard | |
| 8 | Spatial-Omni SO-7B | 49.7% | Official leaderboard | ||
| 9 | | 49.4% | Apr 2026 | Official leaderboard | |
| 10 | video-SALMONN2+ 7B | 48.8% | Official leaderboard | ||
| 11 | | 48.8% | Mar 2025 | Official leaderboard | |
| 12 | OmniVideo-7B (VITA1.5) | 47.1% | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 55.3% | Jul 2026 | Official leaderboard | |
| 2 | | 52.6% | Mar 2026 | Official leaderboard | |
| 3 | | 51.7% | Apr 2026 | Official leaderboard | |
| 4 | | 51.2% | Mar 2026 | Official leaderboard | |
| 5 | | 50.1% | Feb 2026 | Official leaderboard | |
| 6 | | 49.4% | Apr 2026 | Official leaderboard | |
| 7 | | 48.8% | Mar 2025 | Official leaderboard |
Rankings by part
We read 7 parts from the benchmark’s paper, each ranking models on one part of the test, such as Audio Grounding, Visual Grounding and Semantic Alignment. Gemini 3.6 Flash leads 4 of them, and 3 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
Audio Grounding
7 models we track have a score here.
- 1
Nemotron 3 Nano Omni 30B-A3B ↑5 53.4%
- 2
Gemini 3.6 Flash ↓1 52.5%
- 3
Qwen3.5-Omni-Plus ↓1 51.2%
- 4
Qwen3.5-Omni-Flash 48.8%
- 5
MiniCPM-o 4.5 48.4%
- 6
Doubao Seed 2.0 Lite Apr 2026 ↓3 48.4%
- 7
Qwen2.5-Omni-7B 48.1%
Visual Grounding
7 models we track have a score here.
- 1
Qwen3.5-Omni-Plus ↑1 94.8%
- 2
Nemotron 3 Nano Omni 30B-A3B ↑4 94.5%
- 3
Gemini 3.6 Flash ↓2 93.6%
- 4
Qwen3.5-Omni-Flash 89.9%
- 5
Doubao Seed 2.0 Lite Apr 2026 ↓2 88.3%
- 6
MiniCPM-o 4.5 ↓1 84.7%
- 7
Qwen2.5-Omni-7B 78.5%
Semantic Alignment
7 models we track have a score here.
- 1
Gemini 3.6 Flash 65.7%
- 2
Doubao Seed 2.0 Lite Apr 2026 ↑1 62%
- 3
Qwen3.5-Omni-Plus ↓1 59.7%
- 4
MiniCPM-o 4.5 ↑1 58.3%
- 5
Qwen3.5-Omni-Flash ↓1 54%
- 6
Qwen2.5-Omni-7B ↑1 51.7%
- 7
Nemotron 3 Nano Omni 30B-A3B ↓1 50.3%
Spatial Alignment
7 models we track have a score here.
- 1
Qwen3.5-Omni-Flash ↑3 56.3%
- 2
Doubao Seed 2.0 Lite Apr 2026 ↑1 52.7%
- 3
Gemini 3.6 Flash ↓2 52%
- 4
Nemotron 3 Nano Omni 30B-A3B ↑2 51%
- 5
Qwen2.5-Omni-7B ↑2 50%
- 6
MiniCPM-o 4.5 ↓1 49%
- 7
Qwen3.5-Omni-Plus ↓5 48%
Joint A2V
7 models we track have a score here.
- 1
Gemini 3.6 Flash 35.4%
- 2
Qwen3.5-Omni-Plus 35.2%
- 3
Qwen2.5-Omni-7B ↑4 33.1%
- 4
MiniCPM-o 4.5 ↑1 32.8%
- 5
Qwen3.5-Omni-Flash ↓1 32.6%
- 6
Doubao Seed 2.0 Lite Apr 2026 ↓3 29.2%
- 7
Nemotron 3 Nano Omni 30B-A3B ↓1 22.7%
Joint V2A
7 models we track have a score here.
- 1
Gemini 3.6 Flash 42.8%
- 2
Doubao Seed 2.0 Lite Apr 2026 ↑1 39.7%
- 3
Qwen2.5-Omni-7B ↑4 35.7%
- 4
Qwen3.5-Omni-Plus ↓2 35.3%
- 5
MiniCPM-o 4.5 33.5%
- 6
Qwen3.5-Omni-Flash ↓2 33.3%
- 7
Nemotron 3 Nano Omni 30B-A3B ↓1 29.5%
Spatial Reasoning
7 models we track have a score here.
- 1
Gemini 3.6 Flash 44.8%
- 2
Qwen2.5-Omni-7B ↑5 44.3%
- 3
Nemotron 3 Nano Omni 30B-A3B ↑3 44.3%
- 4
MiniCPM-o 4.5 ↑1 43.8%
- 5
Qwen3.5-Omni-Plus ↓3 43.8%
- 6
Qwen3.5-Omni-Flash ↓2 43.2%
- 7
Doubao Seed 2.0 Lite Apr 2026 ↓4 41.7%
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The SAVU-Bench paper
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official SAVU-Bench leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
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
9 of 10 checks pass, 1 warning. Look first at the Task access check. We don’t know whether the tasks are public. Every benchmark
Catalog record 1 warning 1 pass
- Warning Task access We don’t know whether the tasks are public.
- Pass Confidence Checked against a primary source.
Scores 4 pass 1 note
- Pass Independent runs Scores from the benchmark’s paper (12 entries).
- Pass Models scored 7 models we track, plus 5 other entries on its leaderboard.
- Pass Newest score The newest is from July 2026, for Gemini 3.6 Flash.
- Note Frontier models None of the top 10 has a score. The strongest tested is Gemini 3.6 Flash (#45).
- Pass Headroom The best is 55.3%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 7 rows for models we track and 5 other entries at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.