Mundo AI has raised $20 million in Series A funding led by GreatPoint Ventures to build data and evaluation infrastructure designed to help artificial intelligence models better understand audio, video and other real-world sensory information. The financing follows a previously unannounced $4 million seed round, bringing total funding to $24 million. Y Combinator, E12 Ventures and Next Frontier Capital also participated in the Series A.
Mundo is building what it calls the data layer for “perceptual intelligence,” based on the idea that improvements in AI reasoning alone will not be enough to produce systems that can understand and interact naturally with the physical and social world.
The company argues that humans interpret considerably more information from an interaction than what appears in a text transcript.
Tone of voice, facial expressions, gestures, timing, background sounds and social context can all alter the meaning of an interaction, creating challenges for AI systems trained primarily around text or data that does not adequately represent those signals.
Mundo develops datasets, evaluations and applied research for frontier AI laboratories and other AI companies across audio, video and emerging modalities.
Its work includes natural speech-to-speech interactions, detailed video understanding and new categories of data where established training and evaluation approaches have not yet been developed.
The company’s datasets and evaluations are already being used by leading AI labs developing multimodal models.
Mundo plans to use the Series A funding to expand its research, engineering and operations teams as it develops additional data and evaluation infrastructure for perceptual AI systems.
The company’s current strategy represents an expansion of its original focus.
Mundo initially emerged from Y Combinator’s Winter 2025 batch with a mission to address the shortage of high-quality multilingual AI training data.
The founders saw a substantial gap between the performance of AI models in English and many other languages, which they attributed partly to limited quantities of high-quality native-language training data.
Mundo developed an approach based on working directly with native speakers and establishing operations in countries where target languages are spoken.
Its platform was built to support data collection, generation, annotation and quality assurance for AI developers requiring multilingual datasets.
The company has since broadened that thesis beyond language.
Its current platform is focused on the wider challenge of teaching models to understand real-world sensory experiences, including how events unfold over time and how audio, visual information and context interact with one another.
Mundo describes one of the central challenges in audio as the gap between recognizing speech and genuinely understanding a conversation.
On the video side, the company focuses on areas where vision systems can miss context, sequencing and physical intent.
For emerging AI modalities, Mundo is targeting research problems where the necessary datasets may not exist yet and need to be designed around the specific capability researchers are attempting to develop.
Mundo was founded by Jason Liao, Naijide Anwaer, Garreth Lee and Kenneth Wu.
The founders bring backgrounds spanning machine learning research, quantitative finance, Amazon Web Services, Binance.US, Cohere and Hugging Face.
The Series A reflects a broader shift in AI development toward multimodal systems that need to understand more than written language.
Mundo’s thesis is that the next major improvements in model capabilities will require not only larger models and additional compute, but new forms of training data and better methods for measuring how accurately AI systems perceive real-world situations.
