MaineCoon has been introduced as the first video model designed to enhance social interactions through features such as facial expressions, emotions, fluid conversation, and audio-lip synchronization. It boasts impressive specifications, operating with 22 billion parameters and achieving 47.5 frames per second on a single H100 GPU, while generating content in real-time at a cost of less than $0.001 per second. This model utilizes an agentic streaming inference framework with three auxiliary models to effectively manage its cache and lookahead buffer. Unlike traditional AI video, which relies on prompts and pre-recorded clips, MaineCoon allows for real-time interaction, transforming the viewing experience into a dynamic presence. Interested users can learn more and apply for early access at mainecoon.tech.
Generative AIVideo ModelReal-Time Interaction+1
Generative AIVideo ModelReal-Time InteractionAI Presence
MaineCoon Launches as First Real-Time Interactive Video Model Focusing on Social Interactions
Twitter· François Chollet· Monday, June 22, 2026
More Articles From This Day
Open SourceGenerative AI+2
Open SourceGenerative AILLMCost Optimization
Open Weights AI Models Achieve SOTA Results, Driving Cost Optimization in AI Applications
Open weights AI models are currently demonstrating state-of-the-art (SOTA) results across various tasks, nearing the frontier in areas such as coding. The ability of open weights to maintain a narrow gap from leading models enables increased value creation in the AI sector. This trend is advantageous for frontier labs as well, as reducing overall task costs can enhance AI usage. Frontier models continue to be utilized for essential functions like planning and orchestration, while the applied layer of AI is positioned to optimize workloads by leveraging more affordable or specialized open models. Additionally, GLM-5.2 has recently outperformed competitors in web design evaluations.