OpenHome: Can Video Models Understand What Really Happens on a Smart-Home Camera?

OpenHome is a multiple-choice video QA benchmark built from real smart-home camera clips. The clips are short, fixed-viewpoint, often recorded at night or in infrared, and come with no narration or subtitles, so models must reason from sparse visual signals alone. The benchmark is organised into three levels that move from Perception through Reasoning to Insight: reading details of a subject and its interactions and deciding what the homeowner should do about it.

Dataset at a glance

Questions	2,254 (four options, A–D)
Clips	1,118 real-world clips, 5.97 hours total, mean 19.2 s
Conditions	~50% day / ~47% night (22% infrared), mostly outdoor
Subjects	animals, vehicles, people and natural phenomena
Split	test only

The three levels cover 10 tasks:

Category	Tasks	Questions
Perception: what appears in the clip	Attribute, Recognition, Change, Count	989
Reasoning: how it relates in cause, space and time	Causality, Spatial, Temporal	721
Insight: what the subject is up to and what the homeowner should do	Intent, Evidence, Response	544

Every question is answerable from the video frames alone. Questions were drafted by VLMs, hardened against text-only shortcuts, and verified by human annotators against the clip. Correct-answer positions are balanced across A–D (best constant-answer baseline: 26.9%).

Fields

Column	Description
question_id	Unique question ID (<video_id>_q<n>)
video_url	Link to the source clip
category	Perception, Reasoning or Insight
type	One of the 10 tasks
question	Question text
option_A … option_D	Answer options
ground_truth	Correct option letter (A–D)

from datasets import load_dataset
ds = load_dataset("wyzelabs/OpenHome", split="test")

Model performance

Category accuracy (%) of selected closed and open-weight models. Closed-source models lead open-weight models on every level.

#	Model	Perception	Reasoning	Insight	Overall
1	Gemini 4 Argon	81.6	74.8	89.2	81.2
2	GPT-6 Astra	76.4	68.9	86.9	76.6
3	Gemini 3.8 Flash	70.9	63.1	79.0	70.4
4	Muse Wonder 1.0	69.9	62.1	73.7	68.3
5	Gemini 3.1 Pro	65.0	59.8	66.7	63.7
6	Gemini 2.5 Pro	63.5	52.3	57.9	58.6
7	Muse-Glimmer-30B	65.3	52.4	51.7	57.9
8	Grok 4.7	59.6	53.5	57.7	57.2
9	Gemini 3.5 Flash-Lite	60.4	48.8	54.0	55.1
10	Gemma 4 31B-it	62.1	51.5	45.4	54.7
11	GPT-4o	56.8	47.3	55.7	53.5
12	Qwen3.5-397B-A17B	56.5	47.2	50.2	52.0
13	GPT-5 mini	59.2	49.1	42.1	51.8
14	Qwen3.8-Flash-Next	55.9	49.9	46.5	51.7
15	Qwen3.5-27B	53.5	45.6	46.0	49.2
16	MiMo-V2.6-Pro-MOPD	50.6	41.7	46.0	46.6
17	MiMo-V2.6-Flash-MOPD	47.3	40.6	39.0	43.2

Evaluation protocol: frames sampled at 1 fps (max 64), ~200K pixels per frame, no audio or metadata, temperature 0 and minimal reasoning. Full results for all 32 models are in the technical report.

Citation

@techreport{openhome2026,
  title  = {OpenHome: Can Video Models Understand What Really Happens on a Smart-Home Camera?},
  author = {Wyze Labs},
  year   = {2026},
  institution = {Wyze Labs}
}