Category Score V3 leaderboard
LMSpeed Best Multilingual Models
Compare multilingual AI models across cross-language understanding, generation, reasoning transfer, and low-resource robustness benchmarks using a dimension-balanced score.
Methodology 3.0Methodology
Current answer
No model currently meets the formal ranking rules. The table has 7 Estimated models and 56 Provisional models. They are useful signals, but they are not formal ranks.
Available leaderboard data
- Models shown
- 63
- Formally ranked models
- 0
- Benchmark columns
- 5
- Dimensions with evidence
- 2/4
How to read the benchmark bars
Each bar compares models only within the same benchmark column. Bar lengths are relative to the models shown here; they are not Category Scores and cannot be compared across benchmark columns.
| Rank | Model | LMSpeed score | Cross-language understanding | Reasoning transfer | Status | Evidence | Updated | |||
|---|---|---|---|---|---|---|---|---|---|---|
| BenchLM Multilingual score47 models | NOVA-637 models | AA Global-MMLU-Lite4 models | INCLUDE3 models | MMLU-ProX11 models | ||||||
| Estimated models — unranked7 | ||||||||||
| — | Qwen3.7 MaxQwen | 59.9±11.7 | — | 59.0 | — | 86.2 | 87.0 | Estimated | 2/4 dimensions · 3 families | Jul 14, 2026 |
| — | Qwen3.5 | 54.8±12.2 | — | 59.1 | — | — | 84.7 | Estimated | 2/4 dimensions · 2 families | Jul 14, 2026 |
| — | Qwen3.6 PlusQwen | 52.4±12.2 | — | 57.9 | — | — | 84.7 | Estimated | 2/4 dimensions · 2 families | Jul 14, 2026 |
| — | Claude Opus 4.5Anthropic | 52±12.2 | — | 56.7 | — | — | 85.7 | Estimated | 2/4 dimensions · 2 families | Jul 14, 2026 |
| — | Qwen3.7 PlusQwen | 51.9±11.7 | — | 58.8 | — | 83.0 | 85.4 | Estimated | 2/4 dimensions · 3 families | Jul 14, 2026 |
| — | Kimi K2.5MoonshotAI | 44.4±12.2 | — | 56.0 | — | — | 82.3 | Estimated | 2/4 dimensions · 2 families | Jul 14, 2026 |
| — | GLM-5Z.ai | 44±12.2 | — | 55.1 | — | — | 83.1 | Estimated | 2/4 dimensions · 2 families | Jul 14, 2026 |
| Provisional models — unranked56 | ||||||||||
| — | GPT-5.4OpenAI | 61.3±16.1 | 100.0 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Claude Fable 5Anthropic | 61.3±16.1 | 100.0 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Claude Opus 4.8Anthropic | 61.2±17.5 | — | — | — | 87.6 | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | GPT-5.3 CodexOpenAI | 60.3±16.1 | 95.2 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | GPT-5.2OpenAI | 60.3±16.1 | 95.2 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Gemini 3.1 ProGoogle | 59.7±17.3 | — | — | 93.2 | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Claude Sonnet 4.6Anthropic | 59±16.1 | 89.6 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Claude Sonnet 4.5Anthropic | 57.8±16.1 | 84.0 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | GPT-5.1OpenAI | 57.8±16.1 | 84.0 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | GPT-5.1 Codex MaxOpenAI | 57.8±16.1 | 84.0 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | GPT-5OpenAI | 57.8±16.1 | 84.0 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | GPT-5.2 CodexOpenAI | 57.8±16.1 | 84.0 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Gemini 2.5 ProGoogle | 54.8±16.1 | 70.0 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Claude Sonnet 4Anthropic | 54.2±16.1 | 67.2 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | DeepSeek V3.2DeepSeek | 54.2±16.1 | 67.2 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | o4 Mini HighOpenAI | 54.2±16.1 | 67.2 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Claude Opus 4.6Anthropic | 54±17.3 | — | — | 92.2 | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Gemini 3 ProGoogle | 54±17.3 | — | — | 92.2 | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | O3OpenAI | 53.6±16.1 | 64.4 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | O3 ProOpenAI | 53.6±16.1 | 64.4 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Claude Opus 4.1Anthropic | 53.6±16.1 | 64.4 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Claude Haiku 4.5Anthropic | 53±16.1 | 61.6 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | GLM-4.7Z.ai | 52.4±16.1 | 58.8 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Claude 3.5 SonnetAnthropic | 52.4±16.1 | 58.8 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Gemini 3 FlashGoogle | 52.4±16.1 | 58.8 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Llama 3.1Meta | 52.4±16.1 | 58.8 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Mistral Large 3 | 51.8±16.1 | 56.0 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | O1OpenAI | 51.8±16.1 | 56.0 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | MiMo-V2-Flash | 51.8±16.1 | 56.0 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | O3 MiniOpenAI | 49.3±16.1 | 44.8 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | GPT-4oOpenAI | 48.7±16.1 | 42.0 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | GPT-4.1 MiniOpenAI | 48.7±16.1 | 42.0 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Claude 3 HaikuAnthropic | 47.5±16.1 | 36.4 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | GPT-4.1OpenAI | 46.9±16.1 | 33.6 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Gemini 2.5 FlashGoogle | 46.9±16.1 | 33.6 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | GPT-4o MiniOpenAI | 46.3±16.1 | 30.8 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Gemini 3.1 Flash LiteGoogle | 46.3±16.1 | 30.8 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Claude 3 OpusAnthropic | 46.3±16.1 | 30.8 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Qwen3.5-27BQwen | 45.4±16.4 | — | — | — | — | 82.2 | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Qwen3.5-122B-A10BQwen | 45.4±16.4 | — | — | — | — | 82.2 | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Gemini 1.5 ProGoogle | 45.1±16.1 | 25.2 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | GPT-4 TurboOpenAI | 44.5±16.1 | 22.4 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Qwen3.5-35B-A3BQwen | 41.6±16.4 | — | — | — | — | 81.0 | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Phi 4Microsoft | 41.4±16.1 | 8.4 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | DeepSeek R1DeepSeek | 41.4±16.1 | 8.4 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | GPT-4.1 NanoOpenAI | 40.8±16.1 | 5.6 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | DeepSeek V3.1DeepSeek | 40.8±16.1 | 5.6 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | gpt-oss-20bOpenAI | 40.8±16.1 | 5.6 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Llama 4 ScoutMeta | 40.2±16.1 | 2.8 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Grok 3 | 40.2±16.1 | 2.8 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Llama 4 MaverickMeta | 40.2±16.1 | 2.8 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | GLM-4.5 AirZ.ai | 39.8±16.1 | 1.0 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | O1 ProOpenAI | 39.8±16.1 | 1.0 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | GLM-4.5Z.ai | 39.8±16.1 | 1.0 | — | — | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | Qwen3 | 36.7±16.4 | — | — | — | — | 79.4 | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
| — | gpt-oss-120bOpenAI | 31.5±17.3 | — | — | 82.8 | — | — | Provisional | 1/4 dimensions · 1 families | Jul 14, 2026 |
What this leaderboard measures
Which AI model is better suited to multilingual work?
This leaderboard covers cross-language understanding, multilingual generation, reasoning transfer, and low-resource language performance. A current run may cover only some of these dimensions.
Four capability dimensions
The four dimensions come from the category blueprint. Available data may cover only some of them. A dimension without evidence is not presented as a verified capability.
Cross-language understanding
4 benchmark columns currently provide evidence for this dimension.
- BenchLM Multilingual score
- NOVA-63
- AA Global-MMLU-Lite
- INCLUDE
Multilingual generation
Available data has no eligible evidence for this dimension, so it does not affect this Category Score.
Reasoning transfer
1 benchmark columns currently provide evidence for this dimension.
- MMLU-ProX
Low-resource robustness
Available data has no eligible evidence for this dimension, so it does not affect this Category Score.
Multilingual tasks this page can help with
- Translation and localization that preserve meaning, tone, and specialist terms.
- Multilingual support that understands varied phrasing and stays in the target language.
- Cross-language research that finds material in one language and answers in another.
How to choose a model with this leaderboard
- Step 1
Check the rating status first
Only Rated models receive a rank. Estimated and Provisional models do not have a formal position.
- Step 2
Review uncertainty and evidence
When scores are close, do not rely on rank alone. Check uncertainty, dimension coverage, and benchmark count.
- Step 3
Test the real task last
A leaderboard cannot replace your own test. Check quality, speed, price, context, and provider limits together.
Test the exact target language. Check dialects, writing systems, cultural phrasing, and specialist vocabulary.
Rating status guide
Rated
Rated means the evidence and overlap rules are met. The model can receive a formal rank.
Estimated
Estimated means there is useful evidence, but it is not enough for a formal rank.
Provisional
Provisional means evidence is limited or dimension and benchmark-family coverage is below the estimated threshold. Use the result only as an early signal.
Benchmarks and evidence sources
Evidence source names and benchmark groups come from the currently available score data. One source may contribute several benchmarks.
BenchLM
AA Global-MMLU-Lite, BenchLM Multilingual score, INCLUDE, MMLU-ProX, and NOVA-63
How the multilingual model ranking is built
LMSpeed combines eligible third-party benchmarks inside four fixed capability dimensions. Rated models meet the evidence and overlap requirements for a formal rank; Estimated and Provisional models remain visible without receiving a rank.
Read the Category Score methodologyLeaderboard limits
Category Scores use the third-party benchmarks currently included by LMSpeed. Tests can use different data, prompts, and scoring rules. The result is not permanent and cannot represent every real task. Test important choices with your own data and workflow.
Frequently asked questions
Which visible model has the highest formal rank now?
There is no formal number one now. The page has 7 Estimated models and 56 Provisional models. They do not have a formal rank and should not be called the winner.
Can I compare scores across different categories?
No. Each category uses different capability dimensions and evidence. A Category Score is comparable only inside the same leaderboard. Review the matching category for each task.
Are Estimated and Provisional models still useful?
They can help you find candidates, but their evidence is not complete enough for a formal rank. Review coverage and uncertainty, then test the model on a real task.
How often does the leaderboard update?
The leaderboard updates after a new completed score run is published. The current run date and methodology version appear above. LMSpeed does not promise a fixed daily or weekly schedule.
Is the number one model always best for me?
No. Your result also depends on speed, price, context length, tool support, region, and provider limits. Use the leaderboard to narrow the field, then run your own test.
Does a high multilingual score mean every language is supported?
No. The score reflects only the language evidence in the current run. Test each language, dialect, writing system, and specialist domain separately. Low-resource languages need real examples.