# Multilingual MMBench ![Category](https://img.shields.io/badge/Category-Multilingual-1f6feb) ![Metric](https://img.shields.io/badge/Metric-%25%20Accuracy-2da44e) ![Saturation](https://img.shields.io/badge/Saturation-Medium-d29922) ![Created](https://img.shields.io/badge/Created-Jul%202023-555) [![Paper](https://img.shields.io/badge/Paper-arXiv%202307.06281-b31b1b)](https://arxiv.org/abs/2307.06281) [![GitHub](https://img.shields.io/badge/GitHub-open--compass%2FMMBench-181717?logo=github&logoColor=white)](https://github.com/open-compass/MMBench) [![Dataset](https://img.shields.io/badge/Dataset-HuggingFace-ff9d00?logo=huggingface&logoColor=white)](https://huggingface.co/datasets/lmms-lab/MMBench) > **Quick answer:** Multilingual MMBench (`MTL_MMBench_DEV`) is the multilingual development split of the MMBench benchmark (Liu et al., 2023) — the same fine-grained VQA ability dimensions and CircularEval protocol as the English MMBench, evaluated across multiple non-English languages. North Micro Vision scores 63.6% as of August 2026. ## At a Glance **What it tests:** MMBench's ~20 fine-grained perception and reasoning ability dimensions, evaluated via questions and answer choices presented in multiple languages rather than English only. **Why it matters:** Most vision-language benchmarks are English-centric; Multilingual MMBench surfaces the real capability gap many models show when the same visual reasoning task is posed in another language, which matters for any product serving a non-English-speaking user base. **Known limitations:** Coverage and translation quality vary by language, and — like base MMBench — some ability dimensions are approaching saturation for larger frontier models. ## What Multilingual MMBench Measures Multilingual MMBench reuses MMBench's question bank and CircularEval scoring methodology (rotating multiple-choice answer order and requiring correctness across all rotations), but presents the questions and answer options in languages other than English. This isolates how much of a model's visual reasoning capability is coupled to English-language pretraining data versus genuinely transferable across languages. For vision-language models trained with a deliberate multilingual data mixture — such as North Micro Vision, which reports a dedicated multilingual capability group in its evaluation — this split is a meaningful complement to MMMB, testing the same underlying fine-grained reasoning skills as MMBench rather than the broader open-ended multilingual VQA format MMMB uses. Models with English-centric supervised fine-tuning data typically show a noticeable score drop moving from base MMBench to its multilingual split; smaller drops indicate more successful multilingual alignment during training. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Multilingual (fine-grained VQA) | | Metric | % accuracy (CircularEval multiple-choice) | | Saturation | Medium | | Created by | Liu et al. | | Source paper | [Liu et al. 2023](https://arxiv.org/abs/2307.06281) | | GitHub | [open-compass/MMBench](https://github.com/open-compass/MMBench) | | Dataset | [lmms-lab/MMBench on HuggingFace](https://huggingface.co/datasets/lmms-lab/MMBench) | ## How Multilingual MMBench Is Scored Scoring follows the same CircularEval protocol as base MMBench — a question counts as correct only if the model answers correctly across every cyclic shuffling of its multiple-choice options — applied to the multilingual development (`MTL_MMBench_DEV`) split. ## State-of-the-Art Results | Rank | Model | Score | Source | Date | |---|---|---|---|---| | 1 | [North Micro Vision Instruct](/models/cohere-labs/north-micro-vision-instruct) | 63.6% | [North Micro Vision launch blog](https://huggingface.co/blog/CohereLabs/meet-north-micro-vision-instruct) | 2026-08 | *Score sourced from Cohere Labs' North Micro Vision Instruct launch announcement, August 2026, evaluated on the `MTL_MMBench_DEV` split.* ## Multilingual MMBench on Benchgen > No Benchgen results yet — [be the first to run Multilingual MMBench](/benchmarks/liu-et-al/multilingual-mmbench). ## Multilingual MMBench vs Other Benchmarks | Benchmark | What it tests | Saturation | |---|---|---| | **Multilingual MMBench** | MMBench's fine-grained VQA dimensions, evaluated multilingually | Medium | | [MMBench](/benchmarks/liu-et-al/mmbench) | The same ability dimensions, English only | Medium | | [MMMB](/benchmarks/sun-et-al/mmmb) | Broader open-ended multilingual multimodal QA across 6 languages | Low | Use Multilingual MMBench alongside base MMBench to directly measure a model's English-to-multilingual capability gap on identical reasoning tasks; use MMMB for a broader, purpose-built multilingual benchmark. ## Run Multilingual MMBench on Your Model Benchgen lets you evaluate your own vision-language model's multilingual visual reasoning and compare it directly against its English-only MMBench score. ## Related Models & Benchmarks - Top performers: [North Micro Vision Instruct](/models/cohere-labs/north-micro-vision-instruct) - Related benchmarks: [MMBench](/benchmarks/liu-et-al/mmbench) · [MMMB](/benchmarks/sun-et-al/mmmb) ## Frequently Asked Questions

What is Multilingual MMBench?

Multilingual MMBench is the multilingual development split of MMBench, testing the same fine-grained VQA ability dimensions using non-English questions and answer choices.

What does a good Multilingual MMBench score look like?

North Micro Vision Instruct reports 63.6% as of August 2026 — a modest drop from its 68.7% English MMBench score, indicating reasonable multilingual transfer.

Who created Multilingual MMBench?

It uses the same benchmark methodology and question bank created by Liu et al. for MMBench; see the [original paper](https://arxiv.org/abs/2307.06281).

Is Multilingual MMBench saturated?

It shows medium saturation, similar to base MMBench, though it remains discriminative for models without dedicated multilingual training.

Rank	Model	Score
1	lfm3-vl-2b	83.6
2	lfm2-5-vl-3b	79.5
3	north-micro-vision-instruct	63.6

Multilingual MMBench — Results

Multilingual MMBench — Evaluations