H , the Paris startup founded by Google alums, made a big splash last summer when, out of the blue, it announced a seed round of $220 million before releasing a single product. Three months later, still without a product, that splash started to look like a catastrophic flood when three of the company’s five co-founders left over “ operational and business disagreements .”
But the company has kept swimming, and today it’s announcing its first product: Runner H, an “agentic” AI aimed at businesses and developers across tasks like quality assurance and process automation. It’s built atop the startup’s own, proprietary “compact” LLM based on just 2 billion parameters.
H has set up a waitlist for Runner H on its site. CEO Charles Kantor said that it will be releasing APIs to those on the list over the coming days to use agents “off the shelf” that have been pre-built by H, as well as for developers to create their own. Access to the API will also come along with access to something called H-Studio to test and manage how these services work.
Initially, using those APIs will be free, and later there will be a payment model introduced.
Even using compact LLMs, building and running AI is not cheap, especially as competition continues to raise money to develop their own products. TechCrunch has also confirmed that H is raising a Series A to build what Kantor describes as part of the second era of AI — with LLM companies like OpenAI being part of the first era.
“We are lucky to be in the position of building our own models,” Kantor said. “But this second era will be as capital intensive as first one.”
(Recall the $230 million H has already raised — it appears to have added another $10 million since announcing it earlier this year — was a mixture of equity and convertible debt. The long list of investors in that round included individuals like Eric Schmidt, Yuri Milner and Xavier Neil; VCs such as Accel and Creandum; and strategic backers like Amazon, Samsung and UiPath.)
Kantor told TechCrunch that H has quietly been working with a handful of customers in areas like e-commerce, banking, insurance and outsourcing, who have been helping it hone the product.
“Everything [in H] is not based on our creativity but customer feedback,” he said.
Runner H will initially focus on three specific use cases: robotic process automation (RPA), quality assurance, and business process outsourcing.
RPA is an area that has existed for years, using basic scripts to automate the most repetitive tasks that humans have had to perform — such as reading forms, checking boxes, and sending files from one place to another. In fact, a lot of RPA has never been built with AI baked in, even after AI started to develop advanced skills. The idea with Runner H is that it will be able to run RPA across forms, sites, and other templates even when they have been modified (something that might have broken previous scripts), and across a much wider range of sources.
Quality assurance can cover a wide range of applications, but Kantor said that one of the most popular so far has been reducing the “maintenance burdens” around website testing — validating page availability, simulating real user actions, or ensuring compatibility across payment methods — in particular when modifications have been made.
BPO is a catch-all area that will cover not just fixing and improving billing processes, but also speeding up how an agent can use and access data from different sources, and more.
There has been a race among foundational AI companies around how many parameters are going into LLMs. (GPT 4 for example has 175 billion parameters.) But Runner H is taking a very different approach with just 2 billion parameters, both for its LLM and for its computer-vision based “VLM.” Kantor’s argument is that this makes them significantly more efficient in terms of cost and operations, key when working on winning and keeping business deals, and H’s own operational costs.
“We are specialists,” he said. “We are building for the agentic era.”
The company also claims that it works: it says that its compact models outperforms Anthropic’s “Computer Use” by 29% (based on WebVoyager benchmarks) as well as models from Mistral and Meta.