MTVQA is a new benchmark for multilingual text-centric visual question answering (TECVQA), featuring high-quality human expert annotations in nine languages: Arabic, Korean, Japanese, Thai, Vietnamese, Russian, French, German, and Italian. It contains 6,778 question-answer pairs across 2,116 images, covering over 20 fine-grained scenarios from documents and natural scenes. The dataset was meticulously collected and annotated by human experts to ensure visual-textual alignment, with a two-round annotation process involving question generation and evaluation. The benchmark evaluates the performance of various multimodal large language models (MLLMs), including Qwen2.5-VL, InternVL-2.5, GPT-4o, GPT-4V, Claude3, and Gemini. Results show that even the top-performing MLLM, InternVL-2.5, scores significantly lower than human performance, highlighting the need for further improvements in multilingual text-centric visual question answering. The dataset provides nuanced multilingual annotations, aiming to set a new standard for benchmarks and foster advancements in multilingual visual text comprehension. MTVQA is released at https://github.com/bytedance/MTVQA.MTVQA is a new benchmark for multilingual text-centric visual question answering (TECVQA), featuring high-quality human expert annotations in nine languages: Arabic, Korean, Japanese, Thai, Vietnamese, Russian, French, German, and Italian. It contains 6,778 question-answer pairs across 2,116 images, covering over 20 fine-grained scenarios from documents and natural scenes. The dataset was meticulously collected and annotated by human experts to ensure visual-textual alignment, with a two-round annotation process involving question generation and evaluation. The benchmark evaluates the performance of various multimodal large language models (MLLMs), including Qwen2.5-VL, InternVL-2.5, GPT-4o, GPT-4V, Claude3, and Gemini. Results show that even the top-performing MLLM, InternVL-2.5, scores significantly lower than human performance, highlighting the need for further improvements in multilingual text-centric visual question answering. The dataset provides nuanced multilingual annotations, aiming to set a new standard for benchmarks and foster advancements in multilingual visual text comprehension. MTVQA is released at https://github.com/bytedance/MTVQA.