UniPat builds the training data and benchmarks that AI developers use to test whether models can perform real work, from operating business software to reasoning through specialized professional problems. The Beijing startup recently closed a financing at a $1 billion pre-money valuation, The Information reported, according to an accessible Dealroom summary. The round’s size and investors were not disclosed.
The company works across model training and evaluation. It pays specialists in law, finance, healthcare and science to produce data used during reinforcement learning, or post-training, and generates additional synthetic data with its own algorithms. Developers can use that material to refine their models and then measure the results against UniPat’s tests. Dealroom says the company sells training and benchmarking data to researchers and businesses.
SaaS-Bench shows what that testing looks like in practice. UniPat places computer-using agents inside 23 deployable software systems and gives them 106 professional tasks spanning six domains. An agent must move information between applications, preserve records and finish an auditable piece of work; UniPat scores both progress through weighted checkpoints and whether every required step was completed.
Other UniPat benchmarks test coding agents against software-engineering tasks, browser agents on live websites and multimodal models on visual reasoning. Some preserve the same workspace while requirements change over multiple turns, while others test whether an agent can implement an upgrade across an existing codebase. The result is closer to a collection of repeatable work environments than a single leaderboard score.
The closed round is separate from financing talks reported days earlier. Dealroom wrote on September 10 that Alibaba was set to lead a $300 million late-stage investment valuing UniPat at $2.5 billion, with Tencent and HSG also participating, but said negotiations were continuing and terms could change. The later report describes a financing that had already closed at a $1 billion pre-money valuation; it does not confirm that the proposed $300 million round has closed.
The Information’s reporting, as relayed by Dealroom, put UniPat’s customer orders at $30 million at the end of May. Orders are not the same as recognized revenue. The report also said major Chinese model developers had purchased data or services from UniPat, its rival Humanlaya or both, without assigning individual buyers to either provider. The operating test is whether UniPat can convert that order book into recurring demand while maintaining credible evaluations for the laboratories buying its data.