MMLU-Redux
Version 1.0
Manually re-annotated subset of MMLU that fixes wrong labels. Multiple-choice accuracy.
Saturated. This first release has 3,000 questions in 30 subjects. The authors now recommend MMLU-Redux 2.0, and cards that write only MMLU-Redux do not say which version.
- Made by
- University of Edinburgh
- Released
- June 2024
- Kind
- Exam questions
- Level
- University
- Scored
- Checked automatically
- Answers
- Multiple choice
- Tasks
- All public
- Measures
- Capability
- Subject
- Mathematics, Physical sciences, Biological sciences, Economics and business, History and archaeology
- Language
- English
- Interaction
- Single turn
- Items from
- Another benchmark
- Refresh
- New versions replace it
- Run by
- Anyone
- Based on
- MMLU
- In views
- Biology and medicine, Law and finance
- Strongest model tested
-
Qwen3.7-Max #49 on the capability index
- Last in a launch post
- June 2026, dropped after it saturated
- Last model added
- June 2026
In the authors’ words
Maybe not. We identify and analyse errors in the popular Massive Multitask Language Understanding (MMLU) benchmark. Even though MMLU is widely adopted, our analysis demonstrates numerous ground truth errors that obscure the true capabilities of LLMs. For example, we find that 57% of the analysed questions in the Virology subset contain errors. To address this issue, we introduce a comprehensive framework for identifying dataset errors using a novel error annotation protocol. Then, we create MMLU-Redux, which is a subset of 5,700 manually re-annotated questions across all 57 MMLU subjects. We estimate that 6.49% of MMLU questions contain errors. Using MMLU-Redux, we demonstrate significant discrepancies with the model performance metrics that were originally reported. Our results strongly advocate for revising MMLU's error-ridden questions to enhance its future utility and reliability as a benchmark. https://huggingface.co/datasets/edinburgh-dawg/mmlu-redux-2.0.
Qwen3.7-Max leads with 95% among models released by June 2026
Best score per model as labs report them in their own cards, drawn striped. Labs pick their own settings, so these compare less cleanly than independent runs.
The newest model with a score here came out in June 2026, and nobody has reported a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 95.0
Qwen3.7-Max
- 94.9
Qwen3.5 397B-A17B
- 94.5
Qwen 3.6 Plus
- 94.5
Qwen3.7-Plus
- 94.4 Kimi K2 Thinking
- 94.2
Qwen3.5-Omni-Plus
- 94.0
Qwen3.5-122B-A10B
- 93.8
Qwen3-235B-A22B-Thinking
- 93.7
DeepSeek-V3.1
- 93.7
Qwen3-VL-235B-A22B
- 93.5
Qwen3.6 27B
- 93.4
DeepSeek-R1
- 93.3
Qwen3.5-35B-A3B
- 93.3
Qwen 3.6 35B-A3B
- 93.2
Qwen3.5-27B
- 92.9
DeepSeek-R1
- 92.8
Qwen3-Max-2026-01-23
- 92.7
Qwen3-235B-A22B
- 92.7 Kimi K2
- 92.5
Qwen3-Next-80B-A3B
- 92.3
EXAONE 4.0 32B
- 91.2
Qwen3-Coder-Next
- 91.1
Qwen3.5-9B
- 90.9
Qwen3-32B
- 90.8
DeepSeek-V4-Pro
- 89.5
Qwen3-30B-A3B
- 89.3
LongCat-Flash-Thinking
- 89.1
DeepSeek-V3
- 88.8
Qwen3.5-4B
- 88.6
Qwen3-14B
- 87.5
Qwen3-8B
- 86.8
Qwen2.5-72B
- 86.6
Qwen3-Omni-30B-A3B
- 86.3
Nemotron-Cascade 2 30B-A3B
- 85.9
Qwen2.5-VL-72B
- 83.9
Qwen2.5-32B
- 83.7
Qwen3-4B
- 81.7
Qwen-Turbo
- 80.7
Qwen2.5-14B-Instruct-1M
- 80.0
Qwen2.5-14B
- 79.6
Qwen3.5-2B
- 75.4
Qwen2.5-7B
- 73.9
Qwen3-1.7B
- 71.0
Qwen2.5-Omni-7B
- 59.5
Qwen3.5-0.8B
- 1
Qwen3.7-Max 95%
- 2
Qwen3.5 397B-A17B 94.9%
- 3
Qwen 3.6 Plus 94.5%
- 4
Qwen3.7-Plus 94.5%
- 5 Kimi K2 Thinking 94.4%
- 6
Qwen3.5-Omni-Plus 94.2%
- 7
Qwen3.5-122B-A10B 94%
- 8
Qwen3-235B-A22B-Thinking Jul 2025 93.8%
- 9
DeepSeek-V3.1 93.7%
- 10
Qwen3-VL-235B-A22B 93.7%
- 11
Qwen3.6 27B 93.5%
- 12
DeepSeek-R1 May 2025 93.4%
- 13
Qwen3.5-35B-A3B 93.3%
- 14
Qwen 3.6 35B-A3B 93.3%
- 15
Qwen3.5-27B 93.2%
- 16
DeepSeek-R1 92.9%
- 17
Qwen3-Max-2026-01-23 92.8%
- 18
Qwen3-235B-A22B 92.7%
- 19 Kimi K2 Jul 2025 92.7%
- 20
Qwen3-Next-80B-A3B 92.5%
- 21
EXAONE 4.0 32B 92.3%
- 22
Qwen3-Coder-Next 91.2%
- 23
Qwen3.5-9B 91.1%
- 24
Qwen3-32B 90.9%
- 25
DeepSeek-V4-Pro 90.8%
- 26
Qwen3-30B-A3B 89.5%
- 27
LongCat-Flash-Thinking 89.3%
- 28
DeepSeek-V3 89.1%
- 29
Qwen3.5-4B 88.8%
- 30
Qwen3-14B 88.6%
- 31
Qwen3-8B 87.5%
- 32
Qwen2.5-72B 86.8%
- 33
Qwen3-Omni-30B-A3B 86.6%
- 34
Nemotron-Cascade 2 30B-A3B 86.3%
- 35
Qwen2.5-VL-72B 85.9%
- 36
Qwen2.5-32B 83.9%
- 37
Qwen3-4B 83.7%
- 38
Qwen-Turbo 81.7%
- 39
Qwen2.5-14B-Instruct-1M 80.7%
- 40
Qwen2.5-14B 80%
- 41
Qwen3.5-2B 79.6%
- 42
Qwen2.5-7B 75.4%
- 43
Qwen3-1.7B 73.9%
- 44
Qwen2.5-Omni-7B 71%
- 45
Qwen3.5-0.8B 59.5%
Source: Labs’ model cards
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 95% | May 2026 | Lab model card | |
| 2 | | 94.9% | Feb 2026 | Lab model card | |
| 3 | | 94.5% | Apr 2026 | Lab model card | |
| 4 | | 94.5% | Jun 2026 | Lab model card | |
| 5 | Kimi K2 Thinking | 94.4% | Nov 2025 | Lab model card | |
| 6 | | 94.2% | Mar 2026 | Lab model card | |
| 7 | | 94% | Feb 2026 | Lab model card | |
| 8 | | 93.8% | Jul 2025 | Lab model card | |
| 9 | | 93.7% | Aug 2025 | Lab model card | |
| 10 | | 93.7% | Sep 2025 | Lab model card | |
| 11 | | 93.5% | Apr 2026 | Lab model card | |
| 12 | | 93.4% | May 2025 | Lab model card | |
| 13 | | 93.3% | Feb 2026 | Lab model card | |
| 14 | | 93.3% | Apr 2026 | Lab model card | |
| 15 | | 93.2% | Feb 2026 | Lab model card | |
| 16 | | 92.9% | Jan 2025 | Lab model card | |
| 17 | | 92.8% | Jan 2026 | Lab model card | |
| 18 | | 92.7% | Apr 2025 | Lab model card | |
| 19 | Kimi K2 Jul 2025 | 92.7% | Jul 2025 | Lab model card | |
| 20 | | 92.5% | Sep 2025 | Lab model card | |
| 21 | | 92.3% | Jul 2025 | Lab model card | |
| 22 | | 91.2% | Feb 2026 | Lab model card | |
| 23 | | 91.1% | Mar 2026 | Lab model card | |
| 24 | | 90.9% | May 2025 | Lab model card | |
| 25 | | 90.8% | Apr 2026 | Lab model card | |
| 26 | | 89.5% | May 2025 | Lab model card | |
| 27 | | 89.3% | Sep 2025 | Lab model card | |
| 28 | | 89.1% | Dec 2024 | Lab model card | |
| 29 | | 88.8% | Mar 2026 | Lab model card | |
| 30 | | 88.6% | May 2025 | Lab model card | |
| 31 | | 87.5% | May 2025 | Lab model card | |
| 32 | | 86.8% | Sep 2024 | Lab model card | |
| 33 | | 86.6% | Sep 2025 | Lab model card | |
| 34 | | 86.3% | Mar 2026 | Lab model card | |
| 35 | | 85.9% | Feb 2025 | Lab model card | |
| 36 | | 83.9% | Dec 2024 | Lab model card | |
| 37 | | 83.7% | May 2025 | Lab model card | |
| 38 | | 81.7% | Dec 2024 | Lab model card | |
| 39 | | 80.7% | Jan 2025 | Lab model card | |
| 40 | | 80% | Dec 2024 | Lab model card | |
| 41 | | 79.6% | Mar 2026 | Lab model card | |
| 42 | | 75.4% | Dec 2024 | Lab model card | |
| 43 | | 73.9% | May 2025 | Lab model card | |
| 44 | | 71% | Mar 2025 | Lab model card | |
| 45 | | 59.5% | Mar 2026 | Lab model card |
8 models have held number one in labs’ own numbers since its release in June 2024
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Labs’ model cards
Qwen’s own number for Qwen2.5-72B passed 85% in September 2024
That was 3 months after its release. By the same age, 11% of the benchmarks from 2024 had passed 85%. All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Labs’ model cards
Labs headline it less often now, last in June 2026
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. Qwen accounts for 25 of the 35 releases that named it since 2024.
| Date | Launch post | Score | As the post writes it |
|---|---|---|---|
| Jun 1, 2026 | 94.5 | MMLU-Redux | |
| May 20, 2026 | 95 | MMLU-Redux | |
| Apr 24, 2026 | 90.8 | MMLU-Redux (EM) | |
| Apr 22, 2026 | 93.5 | MMLU-Redux | |
| Apr 15, 2026 | 93.3 | MMLU-Redux | |
| Apr 2, 2026 | 94.5 | MMLU-Redux | |
| Mar 30, 2026 | 94.2 | MMLU-Redux | |
| Mar 19, 2026 | 86.3 | MMLU-Redux | |
| Mar 2, 2026 | 59.5 | MMLU-Redux (Thinking) | |
| Mar 2, 2026 | 79.6 | MMLU-Redux (Thinking) | |
| Mar 2, 2026 | 88.8 | MMLU-Redux | |
| Mar 2, 2026 | 91.1 | MMLU-Redux | |
| Feb 24, 2026 | 93.3 | MMLU-Redux | |
| Feb 24, 2026 | 93.2 | MMLU-Redux | |
| Feb 24, 2026 | 94 | MMLU-Redux | |
| Feb 16, 2026 | 94.9 | MMLU-Redux | |
| Jan 23, 2026 | 92.8 | MMLU-Redux | |
| Nov 6, 2025 | Kimi K2 Thinking | 94.4 | MMLU-Redux, no tools |
| Sep 23, 2025 | 93.7 | MMLU-Redux | |
| Sep 22, 2025 | 86.6 | MMLU-Redux | |
| Sep 21, 2025 | 89.3 | MMLU-Redux(acc) | |
| Sep 11, 2025 | 92.5 | MMLU-Redux | |
| Aug 21, 2025 | 93.7 | MMLU-Redux (EM) | |
| Jul 25, 2025 | 93.8 | MMLU-Redux | |
| Jul 15, 2025 | 92.3 | MMLU-Redux (reasoning) | |
| Jul 11, 2025 | Kimi K2 | 92.7 | MMLU-Redux, EM |
| May 28, 2025 | 93.4 | MMLU-Redux (EM) | |
| Apr 29, 2025 | 92.7 | MMLU-Redux | |
| Mar 27, 2025 | 71 | MMLU-redux | |
| Jan 20, 2025 | 92.9 | MMLU-Redux (EM) | |
| Dec 26, 2024 | 89.1 | MMLU-Redux (EM) | |
| Sep 19, 2024 | 86.8 | MMLU-redux |
25 system cards report it, from 6 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.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| Apr 17, 2026 | | 94.2 | MMLU-Redux (Table 4) | Capabilities |
| Mar 19, 2026 | | 86.3 | MMLU-Redux (Table 1) | Capabilities |
| Feb 28, 2026 | | 91.2 | MMLU-Redux (Table 8) | Capabilities |
| Nov 26, 2025 | | 93.7 | MMLU-Redux (Table 6) | Capabilities |
| Sep 23, 2025 | | 89.3 | MMLU-Redux(acc) (Table 2) | Capabilities |
| Sep 22, 2025 | | 86.6 | MMLU-Redux, Instruct (Table 4). Thinking 88.8 (Table 5) | Capabilities |
| Jul 21, 2025 | Kimi K2 technical report | 92.7 | MMLU-Redux (EM) (Table 3) | Capabilities |
| Jul 15, 2025 | | 92.3 | MMLU-Redux (Table 3, Reasoning). Non-reasoning 89.8 (Table 4) | Capabilities |
| May 14, 2025 | | 87.5 | MMLU-Redux, thinking mode (Table 17). Non-thinking 79.5 (Table 18) | Capabilities |
| May 14, 2025 | | 83.7 | MMLU-Redux, thinking mode (Table 17). Non-thinking 77.3 (Table 18) | Capabilities |
| May 14, 2025 | | 90.9 | MMLU-Redux, thinking mode (Table 13). Non-thinking 85.7 (Table 14) | Capabilities |
| May 14, 2025 | | 89.5 | MMLU-Redux, thinking mode (Table 15). Non-thinking 84.1 (Table 16) | Capabilities |
| May 14, 2025 | | 92.7 | MMLU-Redux, thinking mode (Table 11). Non-thinking 89.2 (Table 12) | Capabilities |
| May 14, 2025 | | 88.6 | MMLU-Redux, thinking mode (Table 15). Non-thinking 82.0 (Table 16) | Capabilities |
| May 14, 2025 | | 73.9 | MMLU-Redux, thinking mode (Table 19). Non-thinking 64.4 (Table 20) | Capabilities |
| Mar 26, 2025 | | 71 | MMLU-redux (Table 1) | Capabilities |
| Feb 19, 2025 | | 85.9 | MMLU-redux (Table 4) | Capabilities |
| Jan 26, 2025 | | 80.7 | MMLU-redux (Table 6) | Capabilities |
| Jan 20, 2025 | | 92.9 | MMLU-Redux (EM), DeepSeek-R1 (Table 8, same as final column of Table 3) | Capabilities |
| Dec 26, 2024 | | 89.1 | MMLU-Redux (EM), DeepSeek-V3 (Table 6) | Capabilities |
| Dec 19, 2024 | | 75.4 | MMLU-redux, Qwen2.5-7B-Instruct (Table 8) | Capabilities |
| Dec 19, 2024 | | 86.8 | MMLU-redux, Qwen2.5-72B-Instruct (Table 6) | Capabilities |
| Dec 19, 2024 | | 83.9 | MMLU-redux, Qwen2.5-32B-Instruct (Table 7) | Capabilities |
| Dec 19, 2024 | | 80 | MMLU-redux, Qwen2.5-14B-Instruct (Table 7) | Capabilities |
| Dec 19, 2024 | | 81.7 | MMLU-redux, Qwen2.5-Turbo (Table 7) | Capabilities |
Versions
More in Knowledge
- MMLU-Pro In use Harder 10-option version of MMLU with reasoning-focused questions across 14 subjects.
Intern-S2-397B 89.8%
- MMLU In use 57-subject multiple-choice exam questions from elementary to professional level. Scored by accuracy.
Seed1.8 92.3%
- 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%
- BBEH micro average In use Models solve harder replacements for each of the 23 BIG-Bench Hard tasks, 4,520 examples in total. The score is the share of all examples answered correctly.
Gemma 4 31B IT 74.4%
- HLE-Diamond In use Models answer 1,000 cleaned expert questions from Humanity's Last Exam, 500 on reasoning and 500 on knowledge, closed book. Graded on exact answers. Claude Fable 5.2 77.9%
Data audit
All 5 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 2 pass 3 notes
- Note Independent runs Only labs’ own numbers, so no one else has checked them. Scores from launch posts (32 scores) and system cards (25 scores).
- Pass Models scored 45 models we track.
- Pass Newest score The newest is from June 2026, for Qwen3.7-Plus.
- Note Frontier models None of the top 10 has a score, which is fine for a retired test. The strongest tested is Qwen3.7-Max (#49).
- Note Headroom Saturated. The best is 95%, past 85% since September 2024.
Leaderboard import 1 note
- Note Importer No public leaderboard to read.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.