Paris-based startup H launches proprietary agentic AI and automation platform
H launched its automation platform Studio and Runner H, an AI agent that automates web tasks through natural language commands. Notably, Runner H outperforms recent tools like Anthropic's computer use API, while being built on models a fraction of Sonnet 3.5's size.
H, the Paris-based startup focused on working towards AGI founded by Stanford researcher Charles Kantor, and four former DeepMind researchers, recently announced its automation platform Studio and proprietary agentic AI Runner H. The latter's biggest selling point is that it outperforms Anthropic's computer use API on internal evaluations based on the WebVoyager public benchmark, achieving a 67% success rate (vs Claude's 52%).
According to the startup, the secret behind Runner H's performance is the use of proprietary models, which are specialized, orders of magnitude smaller than general-purpose models such as Sonnet 3.5, and consequently more cost-effective. H-VLM, the startup's vision model trained to interact with visual information, only has 3 billion parameters and yet, it outperforms other vision models in its size class, as well as general-purpose models like GPT-4o and Pixtral Large. Similarly, H-LLM, the language model powering Runner H, only has 2 billion parameters and has been trained with a focus on "fundamental programming and high-level decision-making". H-LLM was benchmarked against the likes of Llama 3.2 3B and Ministral 3B in coding and function calling evaluations.
Runner H is available through the Studio platform (in beta waitlist). The Studio currently hosts Runner H and enables developers to design automation workflows using natural language, and review analytics from past runs. H expects to make the platform more widely accessible in the future, allowing anyone to leverage Studio's tools to create "robust automations for complex workflows like end-to-end e-commerce scenarios and testing (from product discovery to order confirmation) and financial services onboarding (pre-filling multi-step verification processes, document uploads, and compliance checks)."
