Vision & Language Benchmark

WeaveBench

WeaveBench - long-horizon real-world benchmark for computer-use agents over hybrid UI (EMNLP 2026). Scores are pass rates on WeaveBench tasks; repo: https://github.com/weavebench/WeaveBench (paper: https://huggingface.co/papers/2606.09426; site: https://weavebench.github.io/).

Category

Vision & Language

Score type

percent

Scale

100%

higher is better

Models tested

2

What is WeaveBench?

WeaveBench is a benchmark for understanding images combined with text. The official description says: "WeaveBench - long-horizon real-world benchmark for computer-use agents over hybrid UI (EMNLP 2026)". Below we describe in simple words what benchmarks of this type generally examine.

What does it test?

WeaveBench measures understanding images combined with text. In detail, it checks:

  • answering questions about images (figures, photos, charts, memes)
  • reading text inside images and connecting it to the question
  • visual reasoning such as counting, locating and comparing objects
  • combining what is seen with general knowledge

How to read the score

The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better.

Official methodology note: WeaveBench - long-horizon real-world benchmark for computer-use agents over hybrid UI (EMNLP 2026). Scores are pass rates on WeaveBench tasks; repo: https://github.com/weavebench/WeaveBench (paper: https://huggingface.co/papers/2606.09426; site: https://weavebench.github.io/).

Frequently asked questions

What is the WeaveBench benchmark?

WeaveBench is a benchmark for understanding images combined with text. The official description says: 'WeaveBench - long-horizon real-world benchmark for computer-use agents over hybrid UI (EMNLP 2026)'. Below we describe in simple words what benchmarks of this type generally examine.

What does WeaveBench test?

WeaveBench mainly measures understanding images combined with text. Concretely, it checks: answering questions about images (figures, photos, charts, memes); reading text inside images and connecting it to the question; visual reasoning such as counting, locating and comparing objects.

How is the WeaveBench score calculated?

The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better. The declared maximum score of this benchmark is 100.0.

Where do I get WeaveBench data and scores?

The official source is https://github.com/weavebench/WeaveBench. The official benchmark page/reference is linked on this page. Every score on Open LLM Radar is also available as open data: download the scores JSON (http://open-llm-radar-api:8000/api/v1/benchmarks/weavebench/scores) or pull it directly from the Open LLM Radar API with any HTTP client.

Model scores 2 models tested

Reset
# Model Score Date Source Verified
1 Sage-S2.1-Flash
Shanghai AI Laboratory
68.4%
— Report Verified
2 GLM-5.3-Flash
Zhipu AI
60.9%
— Report Unverified

What now?

You've seen how models score on WeaveBench — here's what to do next.