Global MMLU Lite
We show this table for reference; we do not rank on it.
Global MMLU Lite is a multilingual version of MMLU with 400 questions per language, half culturally sensitive and half culturally agnostic. This page holds two independent runs: Artificial Analysis's, in the main table, and Kaggle's.
Benchmark score on AA Global-MMLU-Lite — November 11, 2026
We compile the AA Global-MMLU-Lite rows from secondary reports. Gemini 3.1 Pro leads the table at 93.2%, followed by Gemini 3 Flash (92.7%) and Gemini 3 Pro (92.2%). We do not use these results to rank models overall.
Gemini 3.1 Pro
Gemini 3 Flash
Gemini 3 Pro
15 modelsMultilingualCurrentDisplay onlyUpdated November 11, 2026
Benchmark score table (15 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | Gemini 3.1 ProGoogle | 93.2% | Not reported | Closed |
| 2 | Gemini 3 FlashGoogle | 92.7% | Not reported | Closed |
| 3 | Gemini 3 ProGoogle | 92.2% | Not reported | Closed |
| 4 | Claude Opus 4.6 (Adaptive)Anthropic | 92.2% | Not reported | Closed |
| 5 | Claude Opus 4.5 ThinkingAnthropic | 91.3% | Not reported | Closed |
| 6 | GPT-5 (high)OpenAI | 90.7% | Not reported | Closed |
| 7 | GPT-5.1OpenAI | 90.6% | Not reported | Closed |
| 8 | Gemini 2.5 ProGoogle | 90.3% | Not reported | Closed |
| 9 | Grok 4xAI | 89.5% | Not reported | Closed |
| 10 | GPT-5.4 miniOpenAI | 89.4% | Not reported | Closed |
| 11 | Qwen3.5-122B-A10BAlibaba | 89.1% | Not reported | Open |
| 12 | GPT-5 miniOpenAI | 87.4% | Not reported | Closed |
| 13 | Qwen3.5-27BAlibaba | 86.8% | Not reported | Open |
| 14 | Qwen3.5-35B-A3BAlibaba | 86.4% | Not reported | Open |
| 15 | DeepSeek-R1DeepSeek | 86.0% | Not reported | Open |
Global MMLU Lite: Artificial Analysis’s run versus Kaggle’s run
The highest Global MMLU Lite score in Artificial Analysis’s run, the table above, is 93.2 (Gemini 3.1 Pro). The best row in Kaggle’s own run is 95.7 (Claude Opus 5). Only 2 models have both a score from Artificial Analysis and a Kaggle run we can match, too few to line up model by model, so the numbers below compare the boards, not one model under two setups.
Both are independent runs of Global MMLU Lite, and both publish an average across languages. Kaggle’s average covers 16 language dimensions; Artificial Analysis averages the languages on its own board. Two teams ran it on two harnesses, so read a gap as a difference between setups. How Kaggle Benchmarks reruns tests.
| Source | Rows | Top score | Top row | Median |
|---|---|---|---|---|
| Artificial Analysis (table above) | 15 | 93.2 | Gemini 3.1 Pro | 90.3 |
| Kaggle’s run | 45 | 95.7 | Claude Opus 5 | 89.5 |
Artificial Analysis is the score in the table above. Every Kaggle row is in the Kaggle table below. Source: kaggle.com/benchmarks/cohere-labs/global-mmlu-lite.
Global MMLU Lite (Kaggle)
Published proportions are displayed as percentages. Model configurations, evaluation dates, task versions, and source uncertainty remain attached to the snapshot. Missing results are not zero. How Kaggle Benchmarks reruns tests.
| Source rank | Source model or configuration | Kaggle average | Source setup | Parameters (B) | Open / closed |
|---|---|---|---|---|---|
| 1 | Claude Opus 5 | 95.7% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 2 | Gemini 3.5 Flash | 95.4% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 3 | Gemini 4 Argon | 95.4% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 4 | Claude Opus 5.5 | 95.4% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 5 | Gemini 3.8 Flash | 95.3% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 6 | Gemini 3.7 Flash | 94.9% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 7 | Claude Opus 4.8 | 94.9% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 8 | Gemini 3 Pro Preview | 94.5% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 9 | Claude Opus 4.1 | 94.3% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 10 | Gemini 3.6 Flash | 94.3% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 11 | GPT-5.6 Sol | 94.1% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 12 | GPT-6 Sol | 94.1% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 13 | Grok 4.5 | 93.3% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 14 | Gemini 2.5 Pro | 93.2% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 15 | GPT-5.6 Terra | 92.7% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 16 | Claude Sonnet 5.5 | 92.6% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 17 | GPT-6 Astra | 92.6% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 18 | GPT-6 Luna | 92.0% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 19 | Gemini 2.5 Flash | 91.5% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 20 | Gemini 2.5 Flash Preview | 90.9% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 21 | Gemma 4 31B | 90.7% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 22 | Claude Sonnet 4 | 90.6% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 23 | Gemini 3.5 Flash-Lite | 89.5% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 24 | GPT-5 | 89.0% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 25 | Grok 4 | 88.8% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 26 | Qwen 3 235B A22B Instruct 2506 | 88.0% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 27 | GPT-4.1 | 87.5% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 28 | Claude Haiku 5.5 | 87.3% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 29 | o4 mini | 87.0% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 30 | Gemma 4 26B A4B | 85.7% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 31 | Command A | 83.9% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 32 | GPT-5.6 Luna | 83.1% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 33 | Claude 3.7 Sonnet | 80.8% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 34 | deepseek-v3.1 | 80.4% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 35 | GPT-6.1 Sol | 78.9% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 36 | Mistral Small 3.1 | 78.5% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 37 | o3 mini | 78.0% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 38 | Gemma 3 27B | 76.3% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 39 | Granite 4.0 Small | 75.0% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 40 | Aya Expanse 32B | 73.5% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 41 | DeepSeek-R1 | 67.4% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 42 | Grok 3 Mini | 67.3% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 43 | Gemma 3 4B | 65.1% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 44 | Claude 3.5 Haiku | 61.1% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
| 45 | Mistral Medium 3 | 55.1% | Cohere Labs on KaggleOriginal result | Not reported | Unknown |
Among the reported AA Global-MMLU-Lite rows, Gemini 3.1 Pro is first at 93.2%. The third row is 1.0 points behind. The broader top-10 range is 3.8 points, so many of the published results sit in a relatively narrow band.
15 models have been evaluated on AA Global-MMLU-Lite. The benchmark falls in the Multilingual category. AA Global-MMLU-Lite is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About AA Global-MMLU-Lite
Year
2026
Tasks
Multilingual knowledge questions
Format
Accuracy
Difficulty
Multilingual professional knowledge
The main table is the average across languages that Artificial Analysis publishes, stored as a display-only multilingual score. Kaggle's run of the same test, an average of 16 language dimensions, is shown in its own table with a comparison of the two.
Freshness and provenance
Version
AA Global-MMLU-Lite 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does AA Global-MMLU-Lite measure?
Global MMLU Lite is a multilingual version of MMLU with 400 questions per language, half culturally sensitive and half culturally agnostic. This page holds two independent runs: Artificial Analysis's, in the main table, and Kaggle's.
Which model scores highest on AA Global-MMLU-Lite?
Gemini 3.1 Pro by Google currently leads with a score of 93.2% on AA Global-MMLU-Lite.
How many models are evaluated on AA Global-MMLU-Lite?
15 AI models have published results on AA Global-MMLU-Lite in the BenchLM catalog.
Compare top models on AA Global-MMLU-Lite
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 7,000+ subscribers.
One email each week. Unsubscribe anytime.