H, a Parisian startup founded by Google alumni, made a splash last summer when it suddenly announced a $220 million seed round before releasing a single product. Three months later, there was still no product, but the splash was a devastating flood as three of the company's five co-founders left leaving behind “operational and business differences.” It started to look like.
But the company kept swimming and today announced its first product, the Runner H. This is an “agent” AI aimed at businesses and developers across tasks such as quality assurance and process automation. It is built on top of the startup's proprietary “compact” LLM, which is based on just 2 billion parameters.
H has set up a waiting list for Runner H on the site. CEO Charles Cantor will release an API to those on the list in the coming days for using Company H's pre-built “off-the-shelf” agents, as well as for developers to create their own. He said he plans to do so. Access to the API also includes access to something called H-Studio for testing and managing how these services work.
Initially, these APIs will be free to use, but a payment model will be introduced later.
Even with compact LLMs, building and running AI isn't cheap, especially as competition continues to attract funding to develop unique products. TechCrunch also confirmed that H is raising Series A funding to build what Kanter describes as part of the second era of AI. LLM companies like OpenAI are part of the first era.
“We are fortunate to be in a position where we can build our own model,” Kanter said. “But this second era will be just as capital-intensive as the first.”
(Recall that the $230 million that H has already raised, with an additional $10 million apparently added since the announcement earlier this year) was a combination of equity and convertible debt. The long list of investors in that round also included individuals like Eric Schmidt and Yuri Milner. , Xavier Niel, VCs such as Accel and Creandum, and strategic backers such as Amazon, Samsung, and UiPath.
Kanter told TechCrunch that H is quietly working with a small number of customers in areas such as e-commerce, banking, insurance and outsourcing who are helping hone the product.
“all [in H] “It's not based on our creativity, it's based on 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 a field that has been around for many years, using basic scripts to do most things that humans had to do, such as reading forms, checking boxes, and sending files from one location to another. Automate repetitive tasks. In fact, many RPAs were never built with AI, even after AI began to develop advanced skills. The idea of Runner H is to allow RPA to run across a wider range of sources, even if forms, sites, and other templates have changed (previous scripts may have been broken).
Quality assurance includes a wide range of uses, but one of the most popular by far is website testing (verifying page availability, simulating real user actions, compatibility between payment methods, etc.). Kanter says the goal is to reduce the “maintenance burden” associated with securing Especially if changes are made.
BPO is a comprehensive area that covers not only fixing and improving the billing process, but also speeding up the way agents use and access data from various sources.
There has been a competition among basic AI companies over how many parameters to put into LLM. (GPT 4, for example, has 175 billion parameters.) But Runner H takes a completely different approach, using just 2 billion parameters for both LLM and computer vision-based “VLM.” I am. Mr. Kanter's argument is that this will result in significant cost and operational efficiencies, which are important when trying to win and retain business, and for Company H's own operational costs.
“We are experts,” he said. “We are building towards the agent era.”
The company also claims that it works well. The company says its compact model outperforms Anthropic's “computer utilization” by 29% (based on the WebVoyager benchmark), similar to models from Mistral and Meta.
TechCrunch has a newsletter focused on AI. Sign up here to get it delivered to your inbox every Wednesday.