BenchMAX
Models follow instructions, write code, reason, use tools, translate, and retrieve information across 17 languages. Ten tasks use rule checks, model-judged wins, code tests, answer matching, or translation scores.
The suite translates and adapts existing evaluations, including IFEval, HumanEval+, LiveCodeBench, GPQA, and RULER. Human post-editing checks most translations. Component scores use different protocols and are not a single raw accuracy.
- Made by
- Nanjing University, Shanghai Artificial Intelligence Laboratory, University of Zurich, and Carnegie Mellon University
- Released
- February 2025
- Kind
- Reasoning in many languages
- Level
- University
- Scored
- Graded by a model
- Answers
- Open-ended
- Tasks
- All public
- Measures
- Capability
- Language
- Many languages
- Interaction
- Agentic
- Environment
- APIs and tools
- Items from
- Another benchmark, Translated
- Refresh
- One fixed set
- Run by
- Anyone
- Based on
- GPQA Diamond, HumanEval, HumanEval+, IFEval, LiveCodeBench, RULER
- In views
- Agents, Many languages
- Strongest model tested
-
DeepSeek-V3 #155 on the capability index
- Last model added
- March 2025
- Scores refreshed
- November 20th, 2026
In the authors’ words
Previous multilingual benchmarks focus primarily on simple understanding tasks, but for large language models(LLMs), we emphasize proficiency in instruction following, reasoning, long context understanding, code generation, and so on. However, measuring these advanced capabilities across languages is underexplored. To address the disparity, we introduce BenchMAX, a multi-way multilingual evaluation benchmark that allows for fair comparisons of these important abilities across languages. To maintain high quality, three distinct native-speaking annotators independently annotate each sample within all tasks after the data was machine-translated from English into 16 other languages. Additionally, we present a novel translation challenge stemming from dataset construction. Extensive experiments on BenchMAX reveal varying effectiveness of core capabilities across languages, highlighting performance gaps that cannot be bridged by simply scaling up model size. BenchMAX serves as a comprehensive multilingual evaluation platform, providing a promising test bed to promote the development of multilingual language models. The dataset and code are publicly accessible.
Llama 3.3 70B leads with 85.2 among models released by March 2025 Llama 3.3 70B leads with 85.2 among models released by March 2025
Best score per entry from the benchmark's own leaderboard, read on November 20th, 2026. The authors report no overall score. The headline is rule-based instruction following, the first task in their main table. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.
The newest model with a score here came out in March 2025, and nobody has tested 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
- 85.2
Llama 3.3 70B
- 83.9
DeepSeek-V3
- 80.8 GPT-4o
- 80.8
Qwen2.5-72B
- 79.1 GPT-4o mini
- 78.1
Qwen2.5-32B
- 78.0
DeepSeek-R1-Distill-Llama-70B
- 76.2
Llama 3.1-70B
- 72.6
Gemma 3 4B
- 67.3
DeepSeek-R1-Distill-Qwen-32B
- 65.9
Qwen2.5-7B
- 63.0
Gemma 2 9B
- 62.6
Llama 3.1-8B
- 62.4
Gemma 2 27B
- 61.9
Aya Expanse 32B
- 51.9
InternLM2.5-20B-Chat
- 51.3
InternLM3-8B-Instruct
- 51.2
Aya Expanse 8B
- 51.0
Llama 3.2 3B
- 49.7 DeepSeek-R1-Distill-Llama 8B
- 47.1
Gemma 3 1B
- 46.7
DeepSeek-R1-Distill-Qwen-7B
- 45.7 InternLM2.5 7B
- 34.0
Llama 3.2 1B
- 1
Llama 3.3 70B 85.2
- 2
DeepSeek-V3 83.9
- 3 GPT-4o Aug 2024 80.8
- 4
Qwen2.5-72B 80.8
- 5 GPT-4o mini 79.1
- 6
Qwen2.5-32B 78.1
- 7
DeepSeek-R1-Distill-Llama-70B 78.0
- 8
Llama 3.1-70B 76.2
- 9
Gemma 3 4B 72.6
- 10
DeepSeek-R1-Distill-Qwen-32B 67.3
- 11
Qwen2.5-7B 65.9
- 12
Gemma 2 9B 63.0
- 13
Llama 3.1-8B 62.6
- 14
Gemma 2 27B 62.4
- 15
Aya Expanse 32B 61.9
- 16
InternLM2.5-20B-Chat 51.9
- 17
InternLM3-8B-Instruct 51.3
- 18
Aya Expanse 8B 51.2
- 19
Llama 3.2 3B 51.0
- 20 DeepSeek-R1-Distill-Llama 8B 49.7
- 21
Gemma 3 1B 47.1
- 22
DeepSeek-R1-Distill-Qwen-7B 46.7
- 23 InternLM2.5 7B 45.7
- 24
Llama 3.2 1B 34.0
Source: Official BenchMAX leaderboard
- 85.2
Llama 3.3 70B
- 83.9
DeepSeek-V3
- 80.8 GPT-4o
- 80.8
Qwen2.5-72B
- 79.1 GPT-4o mini
- 78.1
Qwen2.5-32B
- 78.0
DeepSeek-R1-Distill-Llama-70B
- 76.2
Llama 3.1-70B
- 72.6
Gemma 3 4B
- 67.3
DeepSeek-R1-Distill-Qwen-32B
- 65.9
Qwen2.5-7B
- 63.0
Gemma 2 9B
- 62.6
Llama 3.1-8B
- 62.4
Gemma 2 27B
- 61.9
Aya Expanse 32B
- 51.9
InternLM2.5-20B-Chat
- 51.3
InternLM3-8B-Instruct
- 51.2
Aya Expanse 8B
- 51.0
Llama 3.2 3B
- 47.1
Gemma 3 1B
- 46.7
DeepSeek-R1-Distill-Qwen-7B
- 34.0
Llama 3.2 1B
- 1
Llama 3.3 70B 85.2
- 2
DeepSeek-V3 83.9
- 3 GPT-4o Aug 2024 80.8
- 4
Qwen2.5-72B 80.8
- 5 GPT-4o mini 79.1
- 6
Qwen2.5-32B 78.1
- 7
DeepSeek-R1-Distill-Llama-70B 78.0
- 8
Llama 3.1-70B 76.2
- 9
Gemma 3 4B 72.6
- 10
DeepSeek-R1-Distill-Qwen-32B 67.3
- 11
Qwen2.5-7B 65.9
- 12
Gemma 2 9B 63.0
- 13
Llama 3.1-8B 62.6
- 14
Gemma 2 27B 62.4
- 15
Aya Expanse 32B 61.9
- 16
InternLM2.5-20B-Chat 51.9
- 17
InternLM3-8B-Instruct 51.3
- 18
Aya Expanse 8B 51.2
- 19
Llama 3.2 3B 51.0
- 20
Gemma 3 1B 47.1
- 21
DeepSeek-R1-Distill-Qwen-7B 46.7
- 22
Llama 3.2 1B 34.0
Source: Official BenchMAX leaderboard
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 85.2 | Dec 2024 | Official leaderboard | |
| 2 | | 83.9 | Dec 2024 | Official leaderboard | |
| 3 | GPT-4o Aug 2024 | 80.8 | Aug 2024 | Official leaderboard | |
| 4 | | 80.8 | Sep 2024 | Official leaderboard | |
| 5 | GPT-4o mini | 79.1 | Jul 2024 | Official leaderboard | |
| 6 | | 78.1 | Sep 2024 | Official leaderboard | |
| 7 | | 78.0 | Jan 2025 | Official leaderboard | |
| 8 | | 76.2 | Jul 2024 | Official leaderboard | |
| 9 | | 72.6 | Mar 2025 | Official leaderboard | |
| 10 | | 67.3 | Jan 2025 | Official leaderboard | |
| 11 | | 65.9 | Sep 2024 | Official leaderboard | |
| 12 | | 63.0 | Jun 2024 | Official leaderboard | |
| 13 | | 62.6 | Jul 2024 | Official leaderboard | |
| 14 | | 62.4 | Jun 2024 | Official leaderboard | |
| 15 | | 61.9 | Oct 2024 | Official leaderboard | |
| 16 | | 51.9 | Aug 2024 | Official leaderboard | |
| 17 | | 51.3 | Jan 2025 | Official leaderboard | |
| 18 | | 51.2 | Oct 2024 | Official leaderboard | |
| 19 | | 51.0 | Sep 2024 | Official leaderboard | |
| 20 | DeepSeek-R1-Distill-Llama 8B | 49.7 | Official leaderboard | ||
| 21 | | 47.1 | Mar 2025 | Official leaderboard | |
| 22 | | 46.7 | Jan 2025 | Official leaderboard | |
| 23 | InternLM2.5 7B | 45.7 | Official leaderboard | ||
| 24 | | 34.0 | Sep 2024 | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 85.2 | Dec 2024 | Official leaderboard | |
| 2 | | 83.9 | Dec 2024 | Official leaderboard | |
| 3 | GPT-4o Aug 2024 | 80.8 | Aug 2024 | Official leaderboard | |
| 4 | | 80.8 | Sep 2024 | Official leaderboard | |
| 5 | GPT-4o mini | 79.1 | Jul 2024 | Official leaderboard | |
| 6 | | 78.1 | Sep 2024 | Official leaderboard | |
| 7 | | 78.0 | Jan 2025 | Official leaderboard | |
| 8 | | 76.2 | Jul 2024 | Official leaderboard | |
| 9 | | 72.6 | Mar 2025 | Official leaderboard | |
| 10 | | 67.3 | Jan 2025 | Official leaderboard | |
| 11 | | 65.9 | Sep 2024 | Official leaderboard | |
| 12 | | 63.0 | Jun 2024 | Official leaderboard | |
| 13 | | 62.6 | Jul 2024 | Official leaderboard | |
| 14 | | 62.4 | Jun 2024 | Official leaderboard | |
| 15 | | 61.9 | Oct 2024 | Official leaderboard | |
| 16 | | 51.9 | Aug 2024 | Official leaderboard | |
| 17 | | 51.3 | Jan 2025 | Official leaderboard | |
| 18 | | 51.2 | Oct 2024 | Official leaderboard | |
| 19 | | 51.0 | Sep 2024 | Official leaderboard | |
| 20 | | 47.1 | Mar 2025 | Official leaderboard | |
| 21 | | 46.7 | Jan 2025 | Official leaderboard | |
| 22 | | 34.0 | Sep 2024 | Official leaderboard |
Rankings by part
We read 16 parts from the benchmark’s own leaderboard, each ranking models on one part of the test, such as Rule-based instruction following, Model-based instruction following and Function completion. DeepSeek-V3 leads 9 of them, and 4 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
Rule-based instruction following
22 models we track have a score here.
- 1
Llama 3.3 70B 85.2%
- 2
DeepSeek-V3 83.9%
- 3 GPT-4o Aug 2024 80.8%
- 4
Qwen2.5-72B 80.8%
- 5 GPT-4o mini 79.1%
- 6
Qwen2.5-32B 78.1%
- 7
DeepSeek-R1-Distill-Llama-70B 78%
- 8
Llama 3.1-70B 76.2%
- 9
Gemma 3 4B 72.6%
- 10
DeepSeek-R1-Distill-Qwen-32B 67.3%
Model-based instruction following
17 models we track have a score here.
- 1
DeepSeek-V3 ↑1 59.8%
- 2 GPT-4o Aug 2024 ↑1 50%
- 3
Qwen2.5-72B ↑1 36.9%
- 4
DeepSeek-R1-Distill-Llama-70B ↑3 26.6%
- 5 GPT-4o mini 21.9%
- 6
DeepSeek-R1-Distill-Qwen-32B ↑4 19.2%
- 7
Gemma 2 27B ↑7 18%
- 8
Qwen2.5-32B ↓2 17.3%
- 9
Llama 3.3 70B ↓8 17%
- 10
Llama 3.1-70B ↓2 13.2%
Function completion
21 models we track have a score here.
- 1
DeepSeek-R1-Distill-Llama-70B ↑6 84.6%
- 2
DeepSeek-V3 83.2%
- 3
DeepSeek-R1-Distill-Qwen-32B ↑7 80.6%
- 4 GPT-4o Aug 2024 ↓1 80.1%
- 5 GPT-4o mini 78.7%
- 6
Qwen2.5-72B ↓2 78.6%
- 7
Qwen2.5-32B ↓1 75.8%
- 8
Llama 3.3 70B ↓7 74%
- 9
Llama 3.1-70B ↓1 69.7%
- 10
DeepSeek-R1-Distill-Qwen-7B ↑11 69.3%
Problem solving
20 models we track have a score here.
- 1
DeepSeek-V3 ↑1 60.4%
- 2
DeepSeek-R1-Distill-Llama-70B ↑5 54.8%
- 3
DeepSeek-R1-Distill-Qwen-32B ↑7 54.4%
- 4
Qwen2.5-72B 45.5%
- 5
Qwen2.5-32B ↑1 42.7%
- 6
DeepSeek-R1-Distill-Qwen-7B ↑15 37.3%
- 7 GPT-4o mini ↓2 37%
- 8
Llama 3.3 70B ↓7 34.7%
- 9
Llama 3.1-70B ↓1 29.8%
- 10
Qwen2.5-7B ↑1 24.7%
Math
21 models we track have a score here.
- 1
DeepSeek-V3 ↑1 84.2%
- 2
Llama 3.3 70B ↓1 83.8%
- 3
DeepSeek-R1-Distill-Llama-70B ↑4 82.8%
- 4
Llama 3.1-70B ↑4 79.7%
- 5
Qwen2.5-72B ↓1 77.8%
- 6
Qwen2.5-32B 77.7%
- 7
DeepSeek-R1-Distill-Qwen-32B ↑3 77.3%
- 8 GPT-4o mini ↓3 76.9%
- 9
Gemma 2 27B ↑5 75.3%
- 10
Gemma 2 9B ↑2 72%
Science
21 models we track have a score here.
- 1
DeepSeek-V3 ↑1 47.4%
- 2
DeepSeek-R1-Distill-Llama-70B ↑5 46.1%
- 3 GPT-4o Aug 2024 45.6%
- 4
Llama 3.3 70B ↓3 42.6%
- 5
Qwen2.5-72B ↓1 39.4%
- 6
Qwen2.5-32B 37.7%
- 7
DeepSeek-R1-Distill-Qwen-32B ↑3 37%
- 8
Llama 3.1-70B 35.8%
- 9 GPT-4o mini ↓4 34.1%
- 10
DeepSeek-R1-Distill-Qwen-7B ↑11 28.4%
Question answering
8 models we track have a score here.
- 1
DeepSeek-V3 ↑1 85.2%
- 2 GPT-4o mini ↑3 82.1%
- 3
Qwen2.5-72B ↑1 80.6%
- 4
Qwen2.5-32B ↑2 79.4%
- 5
Llama 3.1-8B ↑8 68.3%
- 6
Llama 3.1-70B ↑2 57.4%
- 7
Qwen2.5-7B ↑4 53.5%
- 8
Llama 3.3 70B ↓7 50.4%
Multiple functions
16 models we track have a score here.
- 1 GPT-4o mini ↑4 70.9%
- 2
DeepSeek-V3 69.2%
- 3
Qwen2.5-32B ↑3 66.7%
- 4
Gemma 2 27B ↑10 64.7%
- 5
DeepSeek-R1-Distill-Llama-70B ↑2 62.1%
- 6
Qwen2.5-72B ↓2 61.8%
- 7
Gemma 2 9B ↑5 61.4%
- 8
DeepSeek-R1-Distill-Qwen-32B ↑2 60.4%
- 9
Aya Expanse 32B ↑6 59.8%
- 10
Qwen2.5-7B ↑1 48.9%
General translation: En to X
16 models we track have a score here.
- 1
DeepSeek-V3 ↑1 33.9
- 2
Llama 3.3 70B ↓1 31.5
- 3
Llama 3.1-70B ↑5 31.1
- 4
Gemma 2 27B ↑10 30.4
- 5 GPT-4o mini 30.3
- 6
Gemma 2 9B ↑6 27.2
- 7
DeepSeek-R1-Distill-Llama-70B 26.0
- 8
Qwen2.5-72B ↓4 25.8
- 9
Aya Expanse 32B ↑6 25.2
- 10
Llama 3.1-8B ↑3 24.6
General translation: X to En
16 models we track have a score here.
- 1
Llama 3.1-70B ↑7 35.1
- 2
Gemma 2 27B ↑12 34.5
- 3
DeepSeek-V3 ↓1 34.5
- 4 GPT-4o mini ↑1 33.9
- 5
Llama 3.3 70B ↓4 33.6
- 6
Qwen2.5-72B ↓2 33.3
- 7
Gemma 2 9B ↑5 33.2
- 8
DeepSeek-R1-Distill-Llama-70B ↓1 33.0
- 9
Aya Expanse 32B ↑6 32.8
- 10
Qwen2.5-32B ↓4 30.5
Domain translation: En to X spBLEU
16 models we track have a score here.
- 1
DeepSeek-V3 ↑1 70.3
- 2 GPT-4o mini ↑3 67.7
- 3
Gemma 2 27B ↑11 64.8
- 4
Llama 3.1-70B ↑4 64.5
- 5
Llama 3.3 70B ↓4 63.5
- 6
Qwen2.5-72B ↓2 60.4
- 7
Gemma 2 9B ↑5 57.5
- 8
Aya Expanse 32B ↑7 54.8
- 9
Qwen2.5-32B ↓3 54.2
- 10
Llama 3.1-8B ↑3 53.9
Domain translation: X to En spBLEU
16 models we track have a score here.
- 1
Llama 3.1-70B ↑7 68.2
- 2
DeepSeek-V3 67.8
- 3 GPT-4o mini ↑2 67.6
- 4
Qwen2.5-72B 66.9
- 5
Gemma 2 27B ↑9 66.2
- 6
Qwen2.5-32B 65.4
- 7
Llama 3.3 70B ↓6 65.0
- 8
Llama 3.1-8B ↑5 62.9
- 9
Aya Expanse 32B ↑6 62.3
- 10
Gemma 2 9B ↑2 61.9
Domain translation: En to X TER
11 models we track have a score here.
- 1 GPT-4o mini ↑4 28.8
- 2
DeepSeek-V3 29.5
- 3
Gemma 2 27B ↑11 34.7
- 4
Llama 3.1-70B ↑4 39.1
- 5
Aya Expanse 32B ↑10 39.3
- 6
Qwen2.5-72B ↓2 41.4
- 7
Gemma 2 9B ↑5 44.3
- 8
Llama 3.1-8B ↑5 45.5
- 9
Aya Expanse 8B ↑9 51.3
- 10
Qwen2.5-7B ↑1 58.2
Domain translation: X to En TER
11 models we track have a score here.
- 1 GPT-4o mini ↑4 23.3
- 2
Llama 3.1-70B ↑6 23.9
- 3
DeepSeek-V3 ↓1 24.0
- 4
Qwen2.5-72B 25.8
- 5
Gemma 2 27B ↑9 26.6
- 6
Llama 3.1-8B ↑7 28.3
- 7
Aya Expanse 32B ↑8 28.6
- 8
Qwen2.5-7B ↑3 31.8
- 9
Gemma 2 9B ↑3 32.5
- 10
Aya Expanse 8B ↑8 40.4
Domain translation: En to X xComet
11 models we track have a score here.
- 1
DeepSeek-V3 ↑1 73.5
- 2 GPT-4o mini ↑3 69.6
- 3
Llama 3.1-70B ↑5 69.4
- 4
Gemma 2 27B ↑10 69.2
- 5
Gemma 2 9B ↑7 65.5
- 6
Qwen2.5-72B ↓2 65.3
- 7
Llama 3.1-8B ↑6 63.1
- 8
Aya Expanse 32B ↑7 62.5
- 9
Aya Expanse 8B ↑9 55.8
- 10
Qwen2.5-7B ↑1 54.4
Domain translation: X to En xComet
11 models we track have a score here.
- 1
DeepSeek-V3 ↑1 80.0
- 2 GPT-4o mini ↑3 79.4
- 3
Llama 3.1-70B ↑5 79.1
- 4
Qwen2.5-72B 78.4
- 5
Gemma 2 27B ↑9 78.2
- 6
Gemma 2 9B ↑6 77.2
- 7
Llama 3.1-8B ↑6 76.6
- 8
Aya Expanse 32B ↑7 76.5
- 9
Qwen2.5-7B ↑2 73.5
- 10
Aya Expanse 8B ↑8 72.0
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official BenchMAX leaderboard
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official BenchMAX leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Multilingual
- MMMLU In use MMLU professionally translated into 14 languages. Multiple-choice accuracy. Claude Mythos Preview 92.7%
- WMT24++ In use English-to-55-language translation of WMT24 test data. Scored with automatic metrics such as MetricX or xCOMET.
K-EXAONE 90.5%
- MMLU-ProX In use MMLU-Pro translated into up to 29 languages with expert review. Multiple-choice accuracy per language.
Qwen3.7-Max 87%
- CLIcK In use Models answer 1,995 multiple-choice questions on Korean language and culture. Questions come from official Korean exams and textbooks in eleven categories.
Solar Open 2 90.7%
- Global MMLU Lite 3.0 In use Models answer 400 human-translated MMLU questions per language in 23 languages. Version 3.0 adds Czech, Hungarian, Oriya, Slovak and Tajik and updates Italian. Accuracy. Claude Opus 5 95.7%
- MMMB In use Models answer multiple-choice questions about images in six languages across 15 categories. The set has about 12,000 questions. LFM2.5-VL-3B 83%
Data audit
8 of 9 checks pass, 1 warning. Look first at the Newest score check. The newest came 20 months ago, in March 2025, for Gemma 3 1B. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 2 pass 2 notes
- Pass Independent runs Scores from the official leaderboard (24 entries).
- Pass Models scored 22 models we track, plus 2 other entries on its leaderboard.
- Warning Newest score The newest came 20 months ago, in March 2025, for Gemma 3 1B.
- Note Frontier models None of the top 10 has a score. The strongest tested is DeepSeek-V3 (#155).
- Note Headroom Scored as a plain number, with no fixed ceiling.
Leaderboard import 3 pass
- Pass Importer We read the board on github.com: 22 rows for models we track and 2 other entries at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
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.