H launches Runner H: a 2B-parameter agentic AI for workflow automation

Raising a massive $220 million seed round without a product is wild. Having three of your five co-founders quit three months later over operational disputes is even wilder. That was the story of Paris-based AI startup H earlier this year.

Now, the startup founded by former Google researchers is finally shipping. On November 20, H launched Runner H, an agentic AI platform designed for enterprise workflow automation and quality assurance.

Small models, big claims

While frontier labs are obsessed with scaling parameters into the hundreds of billions, Runner H is taking the opposite route. The system is powered by a proprietary compact model with just 2 billion parameters, paired with a 2B vision-language model.

Despite the tiny footprint, H claims Runner H outperforms Anthropic’s "Computer Use" by 29% on the WebVoyager benchmark, while also beating models from Meta and Mistral.

CEO Charles Kantor argues that lightweight models are the only economically viable path for agents. Running massive frontier models to click buttons on a screen burns cash fast. A 2B parameter footprint drastically lowers operating expenses for both H and its enterprise customers.

H opened a waitlist on its website and plans to roll out APIs shortly. Users will get access to pre-built agents off the shelf, developer tools to build custom agents, and a management environment called H-Studio. Access will be free at first before transitioning to a paid model.

Fixing broken workflows

Runner H is targeting three specific areas: robotic process automation (RPA), quality assurance (QA), and business process outsourcing (BPO).

Traditional RPA scripts break the moment an engineer updates a website button or changes a form field. Runner H uses vision and reasoning to navigate interfaces even when templates change, making automations far more resilient.

On the QA side, the focus is on testing website availability, mimicking real user behaviors, and verifying payment integrations. For BPO, it handles billing fixes and multi-source data extraction.

Kantor confirmed H has quietly worked with enterprise clients across e-commerce, banking, and insurance to refine the software. The startup has also quietly bumped its total funding to $230 million and is already raising a Series A. Kantor noted that while OpenAI defined the first era of AI, the agentic second era will be just as capital-intensive.

Why it matters

For developers and enterprises, Runner H offers a realistic alternative to brute-forcing workflow automation with monster models. Using an expensive LLM to fill out web forms rarely makes unit economic sense. If a 2B parameter specialist model can reliably interact with web interfaces at a fraction of the cost and latency, it becomes much easier to justify deploying AI agents across everyday business operations.

The launch also intensifies the battle over desktop and browser control. Anthropic kicked off the trend with Computer Use via Claude 3.5 Sonnet, proving general-purpose models could drive software. H is betting that specialized, lightweight architectures will win on cost and speed in production environments.

We think targeting RPA and QA is a shrewd commercial move because enterprise budgets for fixing broken scripts already exist. However, the pressure on reliability will be intense. In billing and banking tasks, a 95% success rate is still a failure. With most of its original founding team gone, H still has to prove that its small-model architecture can handle messy real-world edge cases beyond synthetic benchmarks.