On August 23rd, according to the South China Morning News, the family got OpenAI AMERICAN ARTIFICIAL INTELLIGENCE INITIATIVE, BASED ON CHINA ' S AI LABORATORYDark Side of the Moon(Moonshot AI) Kimi K3 built its first internal model. This highlights the fact that, as AI model development costs are soaring, more and more Western technology companies are moving towards China’s open weight model。
Harvey Local Time announced on Thursday that its new model Harvey Tent was trained after the open weight model Kimi K3. The famous investors of Harvey also included Redwood Capital and Andreasen Horowitz。
It stated that the system had achieved “industry-led” performance in complex legal work。
This means a major shift for Harvey. The company serves large international law firms and corporate clients, and has previously focused on legal applications, customizing closed-source proprietary models of American AI giants such as Anthropic, OpenAI and Google。
AI Policy Researcher Simon Hedlin stated in social media on Friday that Harvey's shift was an excellent example of the value of the open weight model. Developers can use such models for post-training based on data from a particular industry or enterprise to improve accuracy and reduce the cost of reasoning。
Post-training refers to the further optimization of generic basic models using specialized data sets to enable them to be better equipped to perform specific tasks。
Hedlin said: “We are still at a very early stage in exploring what a high-capacity open weight model can achieve.” He added: “Unfortunately, the United States is lagging behind in the development of the front-line open weight model.”
Harvey was founded in the summer of 2022 and obtained an estimate of $11.1 billion in the March round of financing (note: the current exchange rate is about $7,417,400 million)。
In a blog post published on Thursday, Harvey stated that the focus of the company's research over the past six months had been to build “front-line legal intelligence” using an open weight model and to enable law firms to build and deploy specialized models。
According to Harvey, after a comprehensive legal data set trainingHarvey Tenet performed better than its bottom base model and front-line AI systems in the United States, including Fable 5 and GPT-5.6 Sol, in a series of complex, long-cycle legal intelligence missions。
Harvey states that this approach also increases cost efficiency. It states that the open weight model itself usually has a lower cost per Token, and Harvey further reduced costs by reducing the amount of Token consumed in the reasoning process。
Harvey claims that Harvey Tent was trained for two months, using about 150 British dollars B300 GPU。
Meanwhile, according to The Information, on Thursday, citing the American telecommunications giant, AT&T, Vice President Mark Austin, including the British Nemotron seriesOpen Source Model, AN AI QUERY INITIATED BY AT&T EMPLOYEES HAS BEEN PROCESSED。