STEM & Reasoning Benchmark

HLE (text-only)

Humanity's Last Exam, text-only question subset (no images), denoted with a dagger (†) on DeepSeek model cards. Accuracy percent over the text-only subset of the cais/hle dataset (https://huggingface.co/datasets/cais/hle).

Category

STEM & Reasoning

Score type

percent

Scale

100%

higher is better

Models tested

3

What is HLE (text-only)?

Humanity's Last Exam (HLE) is a question set of over 2,500 expert questions from about a hundred domains, designed to be the very edge of human knowledge — created specifically because frontier models had become too good at all existing benchmarks. The score reflects the accuracy of the final answer.

What does it test?

HLE (text-only) measures hard reasoning in science, technology, engineering and math. In detail, it checks:

  • extremely difficult interdisciplinary questions
  • genuine expert reasoning instead of pattern matching
  • compositional thinking across several fields
  • accuracy even when guessing would be plausible

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.

Frequently asked questions

What is the HLE (text-only) benchmark?

Humanity's Last Exam (HLE) is a question set of over 2,500 expert questions from about a hundred domains, designed to be the very edge of human knowledge — created specifically because frontier models had become too good at all existing benchmarks. The score reflects the accuracy of the final answer.

What does HLE (text-only) test?

HLE (text-only) mainly measures hard reasoning in science, technology, engineering and math. Concretely, it checks: extremely difficult interdisciplinary questions; genuine expert reasoning instead of pattern matching; compositional thinking across several fields.

How is the HLE (text-only) 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 HLE (text-only) data and scores?

The official source is https://huggingface.co/datasets/cais/hle. 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/hle-text-only/scores) or pull it directly from the Open LLM Radar API with any HTTP client.

Model scores 3 models tested

Reset
# Model Score Date Source Verified
1 DeepSeek-V4.1-Flash-1017
DeepSeek
44%
17.10.2026 Report Verified
2 GLM-5.3
Zhipu AI
42%
17.10.2026 DeepSeek-V4.1-Flash-1017 model card comparison table (https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017) | † text-only subset of HLE Unverified
3 DeepSeek-V4.1-Flash
DeepSeek
39.1%
17.10.2026 DeepSeek-V4.1-Flash-1017 model card comparison table (https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017) | † text-only subset of HLE Unverified

What now?

You've seen how models score on HLE (text-only) — here's what to do next.