StreamingBench
LLaVA-OneVision 7B 56.4% leads the models we track Seed1.8 84.4% leads since December 2025 Models answer questions while a video streams in, as a human viewer would. It tests real-time visual understanding, omni-source understanding and context understanding.
Qwen reports it as audio-visual interaction.
- Made by
- Lin et al. (arXiv 2411.03628)
- Released
- November 2024
- Kind
- Video understanding
- Level
- Everyday
- Scored
- Checked automatically
- Answers
- Multiple choice
- Tasks
- All public
- Last in a launch post
- September 2026, dropped before it saturated
- Last model added
- September 2026
- Scores refreshed
- November 12th, 2026
In the authors’ words
The rapid development of Multimodal Large Language Models (MLLMs) has expanded their capabilities from image comprehension to video understanding. However, most of these MLLMs focus primarily on offline video comprehension, necessitating extensive processing of all video frames before any queries can be made. This presents a significant gap compared to the human ability to watch, listen, think, and respond to streaming inputs in real time, highlighting the limitations of current MLLMs. In this paper, we introduce StreamingBench, the first comprehensive benchmark designed to evaluate the streaming video understanding capabilities of MLLMs. StreamingBench assesses three core aspects of streaming video understanding: (1) real-time visual understanding, (2) omni-source understanding, and (3) contextual understanding. The benchmark consists of 18 tasks, featuring 900 videos and 4,500 human-curated QA pairs. Each video features five questions presented at different time points to simulate a continuous streaming scenario. We conduct experiments on StreamingBench with 13 open-source and proprietary MLLMs and find that even the most advanced proprietary MLLMs like Gemini 1.5 Pro and GPT-4o perform significantly below human-level streaming video understanding capabilities. We hope our work can facilitate further advancements for MLLMs, empowering them to approach human-level video comprehension and interaction in more realistic scenarios.
Gemini 1.5 pro (snapshot unspecified) tops the leaderboard with 67.1%
Best score per entry from the benchmark's own leaderboard, read on November 12th, 2026. 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. The best model we track is LLaVA-OneVision 7B, at 56.4%.
Seed1.8 leads with 84.4%, 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.
The newest model with a score here came out in August 2024, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- Claude Fable 5.2
-
Gemini 4 Carbon - GPT-6 Astra
- Claude Opus 5.5
- and 16 more
- 67.1 Gemini 1.5 pro (snapshot unspecified)
- 60.1 GPT-4o (snapshot unspecified)
- 57.7 Claude 3.5 Sonnet (snapshot unspecified)
- 56.4 LLaVA-OneVision 7B
- 54.1 Qwen2-VL-7B-Instruct
- 53.9 MiniCPM-V 2.6-8B
- 52.8 LLaVA-NeXT-Video-32B
- 51.4 InternVL-V2-8B
- 51.1 Kangaroo-7B
- 48.7 LongVA-7B
- 43.2
Llama-3-VILA1.5-8B - 42.5 Video-CCAM-14B
- 40.4 VideoLLAMA2-7B
- 1 Gemini 1.5 pro (snapshot unspecified) 1 fps 67.1%
- 2 GPT-4o (snapshot unspecified) 64 frames 60.1%
- 3 Claude 3.5 Sonnet (snapshot unspecified) 20 frames 57.7%
- 4 LLaVA-OneVision 7B 32 frames 56.4%
- 5 Qwen2-VL-7B-Instruct 0.2-1 fps 54.1%
- 6 MiniCPM-V 2.6-8B 32 frames 53.9%
- 7 LLaVA-NeXT-Video-32B 64 frames 52.8%
- 8 InternVL-V2-8B 16 frames 51.4%
- 9 Kangaroo-7B 64 frames 51.1%
- 10 LongVA-7B 128 frames 48.7%
- 11 Llama-3-VILA1.5-8B 14 frames 43.2%
- 12 Video-CCAM-14B 96 frames 42.5%
- 13 VideoLLAMA2-7B 32 frames 40.4%
Source: Official StreamingBench leaderboard
- 1 Seed1.8 84.4%
- 2
Qwen3.8-Omni-Flash 80.8% - 3 Seed1.5-VL 68%
- 4 LLaVA-OneVision 7B 32 frames 56.4%
- 5 Qwen2-VL-7B-Instruct 0.2-1 fps 54.1%
- 6 LongVA-7B 128 frames 48.7%
- 7 Llama-3-VILA1.5-8B 14 frames 43.2%
- 8 VideoLLAMA2-7B 32 frames 40.4%
Sources: Labs’ model cards and official StreamingBench leaderboard
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Gemini 1.5 pro (snapshot unspecified) | 67.1% | 1 fps | Official leaderboard | |
| 2 | GPT-4o (snapshot unspecified) | 60.1% | 64 frames | Official leaderboard | |
| 3 | Claude 3.5 Sonnet (snapshot unspecified) | 57.7% | 20 frames | Official leaderboard | |
| 4 | LLaVA-OneVision 7B | 56.4% | 32 frames | Aug 2024 | Official leaderboard |
| 5 | Qwen2-VL-7B-Instruct | 54.1% | 0.2-1 fps | Aug 2024 | Official leaderboard |
| 6 | MiniCPM-V 2.6-8B | 53.9% | 32 frames | Official leaderboard | |
| 7 | LLaVA-NeXT-Video-32B | 52.8% | 64 frames | Official leaderboard | |
| 8 | InternVL-V2-8B | 51.4% | 16 frames | Official leaderboard | |
| 9 | Kangaroo-7B | 51.1% | 64 frames | Official leaderboard | |
| 10 | LongVA-7B | 48.7% | 128 frames | Jun 2024 | Official leaderboard |
| 11 | Llama-3-VILA1.5-8B | 43.2% | 14 frames | Apr 2024 | Official leaderboard |
| 12 | Video-CCAM-14B | 42.5% | 96 frames | Official leaderboard | |
| 13 | VideoLLAMA2-7B | 40.4% | 32 frames | Jun 2024 | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Seed1.8 | 84.4% | Dec 2025 | Lab model card | |
| 2 | Qwen3.8-Omni-Flash | 80.8% | Sep 2026 | Lab model card | |
| 3 | Seed1.5-VL | 68% | May 2025 | Lab model card | |
| 4 | LLaVA-OneVision 7B | 56.4% | 32 frames | Aug 2024 | Official leaderboard |
| 5 | Qwen2-VL-7B-Instruct | 54.1% | 0.2-1 fps | Aug 2024 | Official leaderboard |
| 6 | LongVA-7B | 48.7% | 128 frames | Jun 2024 | Official leaderboard |
| 7 | Llama-3-VILA1.5-8B | 43.2% | 14 frames | Apr 2024 | Official leaderboard |
| 8 | VideoLLAMA2-7B | 40.4% | 32 frames | Jun 2024 | Official leaderboard |
Labs report scores for 3 models in their cards. Switch the scores above to rank them together.
Number one over time 3 models have held number one with claims included since its release in November 2024
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 StreamingBench 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 StreamingBench leaderboard and labs’ model cards
Labs headline it less often now, last in September 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.
| Date | Launch post | Score | As the post writes it |
|---|---|---|---|
| Sep 18, 2026 | Qwen3.8-Omni-Flash | 80.8 | StreamingBench Audio-Visual Interaction |
| Dec 18, 2025 | Seed1.8 | 84.4 | StreamingBench (proactive) |
| May 12, 2025 | Seed1.5-VL | 68 | StreamingBench (proactive), thinking 68.0 (non-thinking 82.8) |
3 system cards report it, from 2 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 |
|---|---|---|---|---|
| Sep 22, 2026 | Qwen3.8-Omni-Flash technical report | 80.8 | StreamingBench (Table 5) | Capabilities |
| Mar 21, 2026 | Seed1.8 model card | 84.4 | StreamingBench, proactive (Table 3) | Capabilities |
| May 11, 2025 | Seed1.5-VL technical report | 68 | StreamingBench (proactive), thinking (Table 7). Non-thinking 82.8 | Capabilities |
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%