A US artificial intelligence start-up backed by OpenAI has built its first in-house model on Chinese lab Moonshot AI’s Kimi K3, highlighting a growing shift by Western tech firms towards Chinese open-weight systems amid soaring development costs.
San Francisco-based legal tech provider Harvey, whose high-profile backers also include Sequoia Capital and Andreessen Horowitz, said on Thursday that its new model, Harvey Tenet, was post-trained on top of the open-weight Kimi K3 base.
The company said the system achieved “state-of-the-art” performance in complex legal work.
The release marks a significant departure for Harvey, which serves major international law firms and enterprise clients. The start-up previously focused on customising closed proprietary models from US leaders such as Anthropic, OpenAI and Google for legal applications.
Harvey’s pivot was “a great example of open-weight models” enabling developers to post-train systems on specific industry or corporate data for higher accuracy and lower inference costs, AI policy researcher Simon Hedlin wrote on social media on Friday.
Post-training is the process of refining a general-purpose base model with specialised data sets to excel at specific tasks.
“We’re still only in the very earliest stages of exploring what’s possible to do with highly capable open-weight models,” Hedlin said.
“It’s unfortunate that America is lagging behind in developing frontier open-weight models.”
Founded in the summer of 2022, the start-up reached a valuation of US$11 billion in a financing round in March.
Harvey said in a blog post on Thursday that its research over the past six months focused on building “frontier legal intelligence using open-weight models” and enabling law firms to build and deploy specialised models.
Trained on comprehensive legal data sets, Harvey Tenet outperformed both its underlying base model and US frontier systems – including Fable 5 and GPT-5.6 Sol – across a range of complex, long-horizon legal agentic tasks, according to the company.
The approach also improved cost efficiency, according to Harvey. The firm said that while open-weight models naturally offered lower per token prices, it also worked on reducing the number of tokens consumed in inference.
The training for Harvey Tenet was done over two months using around 150 Nvidia B300 graphics processing units (GPUs), the company said.
Open-source models such as Nvidia’s Nemotron series now account for 40 per cent of US telecommunications giant AT&T’s employee AI queries, according to a Thursday report by The Information, citing AT&T vice-president Mark Austin.
Austin reportedly also said that AT&T was not currently using any Chinese open-weight models.
The company was still “evaluating the potential risks” of using them, he reportedly said, and analysing options from Chinese firms including DeepSeek and Moonshot, the Beijing-based company behind Kimi K3. -- SOUTH CHINA MORNING POST
