Artificial intelligence startup Trillion Labs announced on the 9th that it has developed "gWorld-32B," the world's first mobile world model that learns causal relationships in surrounding environments to simulate future changes.
A world model is technology that allows AI agents to visually and logically imagine and simulate the results of specific actions before executing them, operating on principles similar to how humans understand physical laws and predict outcomes of actions. The gWorld-32B unveiled by Trillion Labs is the world's first world model implemented in a mobile environment, reproducing subsequent screen states in real-time based on users' touch inputs as if simulating them.
While existing world models generated next-screen states in image (pixel) units, causing text blurring and shape distortion, gWorld-32B predicts in the form of "executable web code (HTML·CSS)," improving the structural limitations of previous models. This technology generates code equivalent to precise blueprints rather than sending photos of the next screen, enabling real-time high-definition rendering. This reduced rendering failure rates to below 1% while maintaining high accuracy that keeps text and icons clear without distortion.
Through model optimization technology, it recorded GUI (the visual system through which users communicate with devices via on-screen icons and menus) prediction performance exceeding "Llama-4-402B," a super-large AI model with more than 50 times larger parameter scale. This demonstrates that gWorld-32B is a "practical multimodal" model capable of predicting and generating screen changes and interface operations in actual mobile environments.
On KApps, a Korean benchmark specialized for domestic mobile environments measuring Korean mobile app manipulation performance, it also achieved "zero-shot" (AI performing tasks it encounters for the first time without additional learning of specific data) performance deployable immediately without separate additional training. Based on this performance, gWorld-32B proved its technological competitiveness optimized for Korea's unique mobile app ecosystem and UI structure compared to global general-purpose models.
gWorld-32B is executable AI technology applicable in actual industrial settings including mobile agents, automated app testing, and customer service automation, enabling automation across digital service operations.
Trillion Labs CEO Shin Jae-min said, "Beyond AI that asks and answers, the era of 'executable AI' that acts directly in complex digital and physical environments has arrived. This is the result of focusing on world models before government AI strategic initiatives began in earnest, and we expect it to bring innovation to Korea's AI sovereignty and industrial competitiveness in various areas including robotics and public service automation."
Trillion Labs is an AI model startup founded in 2024 by CEO Shin Jae-min, who was a core developer of Naver's HyperCLOVA X. Within one year of establishment, it developed and released a 70B-scale large language model (LLM) from scratch.