Benchmark Almanac

Natural Questions

DeepSeek-V2 (MoE-236B, May 2024) 53.4% leads since May 2024 Dropped

Models answer real Google search questions. Model cards usually use the open-domain version, scored by exact match against short answers from Wikipedia.

The original task reads a full Wikipedia page, while model cards mostly report the closed-book variant.

In the authors’ words

We present the Natural Questions corpus, a question answering data set. Questions consist of real anonymized, aggregated queries issued to the Google search engine. An annotator is presented with a question along with a Wikipedia page from the top 5 search results, and annotates a long answer (typically a paragraph) and a short answer (one or more entities) if present on the page, or marks null if no long/short answer is present. The public release consists of 307,373 training examples with single annotations; 7,830 examples with 5-way annotations for development data; and a further 7,842 examples with 5-way annotated sequestered as test data. We present experiments validating quality of the data. We also describe analysis of 25-way annotations on 302 examples, giving insights into human variability on the annotation task. We introduce robust metrics for the purposes of evaluating question answering systems; demonstrate high human upper bounds on these metrics; and establish baseline results using competitive methods drawn from related literature.

Abstract of Natural Questions: A Benchmark for Question Answering Research (ACL Q19-1026, CC BY 4.0)
Made by Google Released Jan 2019 Trivia and open questions All tasks public Last in a launch post Jun 2025 , dropped before it saturated Last model added Aug 2025
Level
Everyday Most adults, without training
Scored
Checked automatically A program checks the result against an answer key, unit tests or the expected final state.
Answers
Short answer A number, word or short string matched to a key.
google-research-datasets/natural-questions 1,137 stars Paper Dataset Leaderboard ai.google.com
20
Models scored in lab cards
8
Launch posts list it, from 3 labs
18
System cards report it, from 5 labs

DeepSeek-V2 (MoE-236B, May 2024) leads with 53.4% among models released by Aug 2025

Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.

The newest model with a score here came out in Aug 2025, and nobody has reported a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

20 models. Scroll sideways for the rest.

  1. 53.4 DeepSeek-V2 (MoE-236B, May 2024)
  2. 50.4 InternVL3.5-241B-A28B
  3. 47.0 DeepSeek LLM 67B
  4. 40.0 DeepSeek-V3
  5. 39.5 Llama 2-70B
  6. 37.5 PaLM 2-L
  7. 36.1 Gemma 3 27B
  8. 34.5 Gemma 2 27B
  9. 32.8 Llama 2-34B
  10. 31.4 Gemma 3 12B
  11. 31.2 Llama 2-13B
  12. 30.6 Mixtral 8x7B
  13. 29.2 Gemma 2 9B
  14. 28.8 Mistral 7B v0.1
  15. 25.7 Llama 2-7B
  16. 23.0 Gemma 7B
  17. 20.9 Gemma 3n E4B
  18. 20.0 Gemma 3 4B
  19. 12.5 Gemma 2B
  20. 9.48 Gemma 3 1B
  1. 1 DeepSeek-V2 (MoE-236B, May 2024) 53.4%
  2. 2 InternVL3.5-241B-A28B 50.4%
  3. 3 DeepSeek LLM 67B 47%
  4. 4 DeepSeek-V3 40%
  5. 5 Llama 2-70B 39.5%
  6. 6 PaLM 2-L 37.5%
  7. 7 Gemma 3 27B 36.1%
  8. 8 Gemma 2 27B 34.5%
  9. 9 Llama 2-34B 32.8%
  10. 10 Gemma 3 12B 31.4%
  11. 11 Llama 2-13B 31.2%
  12. 12 Mixtral 8x7B 30.6%
  13. 13 Gemma 2 9B 29.2%
  14. 14 Mistral 7B v0.1 28.8%
  15. 15 Llama 2-7B 25.7%
  16. 16 Gemma 7B 23%
  17. 17 Gemma 3n E4B 20.9%
  18. 18 Gemma 3 4B 20%
  19. 19 Gemma 2B 12.5%
  20. 20 Gemma 3 1B 9.48%
#ModelScoreSettingDateSource
1 DeepSeek-V2 (MoE-236B, May 2024) 53.4% May 2024 Model card
2 InternVL3.5-241B-A28B 50.4% Aug 2025 Model card
3 DeepSeek LLM 67B 47% Jan 2024 Model card
4 DeepSeek-V3 40% Dec 2024 Model card
5 Llama 2-70B 39.5% Jul 2023 Model card
6 PaLM 2-L 37.5% May 2023 Model card
7 Gemma 3 27B 36.1% Mar 2025 Model card
8 Gemma 2 27B 34.5% Jun 2024 Model card
9 Llama 2-34B 32.8% Jul 2023 Model card
10 Gemma 3 12B 31.4% Mar 2025 Model card
11 Llama 2-13B 31.2% Jul 2023 Model card
12 Mixtral 8x7B 30.6% Dec 2023 Model card
13 Gemma 2 9B 29.2% Jul 2024 Model card
14 Mistral 7B v0.1 28.8% Sep 2023 Model card
15 Llama 2-7B 25.7% Jul 2023 Model card
16 Gemma 7B 23% Feb 2024 Model card
17 Gemma 3n E4B 20.9% Jun 2025 Model card
18 Gemma 3 4B 20% Mar 2025 Model card
19 Gemma 2B 12.5% Mar 2024 Model card
20 Gemma 3 1B 9.48% Mar 2025 Model card

4 models have held number one since 2023

Lab claims count from the day the card came out, independent runs from the model’s release.

Other labs Meta DeepSeek
Lab claims Not tested since Aug 2025
DeepSeek-V2 (MoE-236B, May 2024) 53.4%

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2023 2024 2025 Near the ceiling (85%) DeepSeek-V2 (MoE-236B, May 2024) DeepSeek LLM 67B Llama 2-70B 0% 20% 40% 60% 80% 100% 2023 2024 2025 Near the ceiling (85%) DeepSeek-V2 (MoE-236B, May 2024) DeepSeek LLM 67B

No launch post has listed it since Jun 2025

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.

Google DeepMind
Mistral AI
DeepSeek
DateLaunch postScoreAs the post writes it
26 Jun 2025 Gemma 3n E4B 20.9 Natural Questions 5-shot (PT)
12 Mar 2025 Gemma 3 27B 36.1 Natural Questions 5-shot (PT)
26 Dec 2024 DeepSeek-V3 40 NaturalQuestions (EM)
27 Jun 2024 Gemma 2 27B 34.5 Natural Questions (5-shot)
21 Feb 2024 Gemma 7B 23 Natural Questions (5-shot)
11 Dec 2023 Mixtral 8x7B 30.6 NQ [detailed table]
27 Sep 2023 Mistral 7B v0.1 28.8 NQ
10 May 2023 PaLM 2-L 37.5 NaturalQuestions (EM), 1-shot

18 system cards report it, from 5 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.

DateDocumentScoreAs the document writes itSection
25 Aug 2025 InternVL3.5-241B-A28B technical report 50.4 NaturalQuestions, OpenCompass (Table 14) Capabilities
17 Jul 2025 Voxtral Small technical report Natural QA 5-shot, text-only, chart without printed numbers (Figure 6) Capabilities
25 Mar 2025 Gemma 3 4B technical report 20 Natural Questions (NQ), PT, 5-shot (Table 9) Capabilities
25 Mar 2025 Gemma 3 27B technical report 36.1 Natural Questions (NQ), PT, 5-shot (Table 9) Capabilities
25 Mar 2025 Gemma 3 1B technical report 9.48 Natural Questions (NQ), PT, 5-shot (Table 9) Capabilities
25 Mar 2025 Gemma 3 12B technical report 31.4 Natural Questions (NQ), PT, 5-shot (Table 9) Capabilities
31 Jul 2024 Gemma 2 9B technical report 29.2 NQ, PT, 5-shot (Table 13) Capabilities
31 Jul 2024 Gemma 2 27B technical report 34.5 NQ, PT, 5-shot (Table 13) Capabilities
7 May 2024 DeepSeek-V2 (MoE-236B, May 2024) technical report 53.4 NaturalQuestions 5-shot, DeepSeek-V2 Chat (RL) (Table 3) Capabilities
13 Mar 2024 Gemma 7B technical report 23 NQ, PT, 5-shot (Table 6) Capabilities
13 Mar 2024 Gemma 2B technical report 12.5 NQ, PT, 5-shot (Table 6) Capabilities
8 Jan 2024 Mixtral 8x7B technical report 30.6 NQ, 5-shot (Table 2) Capabilities
5 Jan 2024 DeepSeek LLM 67B technical report 47 NaturalQuestions, DeepSeek 67B Chat (Table 6) Capabilities
10 Oct 2023 Mistral 7B v0.1 technical report 28.8 NQ, 5-shot (Table 2) Capabilities
19 Jul 2023 Llama 2-7B technical report 25.7 NaturalQuestions, exact match, 5-shot, pretrained (Table 22). 1-shot = 22.7 Capabilities
19 Jul 2023 Llama 2-70B technical report 39.5 NaturalQuestions, exact match, 5-shot, pretrained (Table 22). 1-shot = 33.0 Capabilities
19 Jul 2023 Llama 2-34B technical report 32.8 NaturalQuestions, exact match, 5-shot, pretrained (Table 22). 1-shot = 30.0 Capabilities
19 Jul 2023 Llama 2-13B technical report 31.2 NaturalQuestions, exact match, 5-shot, pretrained (Table 22). 1-shot = 28.0 Capabilities

More in Knowledge

All 109 in Knowledge
  • MMLU-Pro In use Harder 10-option version of MMLU with reasoning-focused questions across 14 subjects. Intern-S2-397B 89.8%
  • SuperGPQA In use 26,529 graduate-level multiple-choice questions across 285 disciplines. Accuracy. Qwen 3.6 Max (Preview) 73.9%
  • WideSearch In use 200 bilingual broad info-seeking tasks: agent collects many atomic facts into a table. Scored by success rate and row/item F1. Tencent Hy4 preview 83.9%
  • KMMLU-Pro In use Models answer multiple-choice questions taken from Korean National Professional Licensure exams. It tests professional knowledge used in Korean industry. Solar Pro 4 79.2%
  • BIG-Bench Extra Hard In use Models solve harder replacements for each of the 23 BIG-Bench Hard tasks, 4,520 examples in total. The score is a harmonic mean of task accuracies. Gemma 4 31B IT 74.4%
  • HLE-Verified In use Verified and revised version of Humanity's Last Exam (1,811 items) with uncertain items removed. Seed2.0 Pro 73.6%