Longbench v2 (≤ 1M)

LongBench v2 (≤1M) is a multiple-choice reading test for very long documents: up to one million tokens of realistic material (books, contract texts, code repositories) with questions that require understanding of the whole document.

Longbench v2 (≤ 1M) measures reading and reasoning over very long inputs. In detail, it checks:

• comprehension of extremely long real-world documents
• combining information from distant sections
• answer correctness instead of keyword matching
• document types in several languages

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

What is the Longbench v2 (≤ 1M) benchmark?

LongBench v2 (≤1M) is a multiple-choice reading test for very long documents: up to one million tokens of realistic material (books, contract texts, code repositories) with questions that require understanding of the whole document.

What does Longbench v2 (≤ 1M) test?

Longbench v2 (≤ 1M) mainly measures reading and reasoning over very long inputs. Concretely, it checks: comprehension of extremely long real-world documents; combining information from distant sections; answer correctness instead of keyword matching.

How is the Longbench v2 (≤ 1M) 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 Longbench v2 (≤ 1M) data and scores?

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/longbench-v2-1m/scores) or pull it directly from the Open LLM Radar API with any HTTP client.