San Francisco legal technology start-up Harvey moves away from proprietary US models, adopting Chinese open-weight Kimi K3 for legal reasoning, signalling a shift towards cost-effective, adaptable AI solutions amid rising training expenses.
Harvey, the San Francisco legal technology start-up backed by OpenAI, has taken a notable step away from the proprietary US models that have dominated its product strategy. According to the company and reporting from the South China Morning Post, its new in-house system, Harvey Tenet, was post-trained on top of Kimi K3, an open-weight model from the Chinese lab Moonshot AI. The move suggests that some Western developers are now looking beyond the leading US model providers as training and inference costs continue to rise.
Harvey said Tenet is its first model built through post-training rather than simple customisation of a closed system. In its announcement, the company said the model is intended for legal reasoning and long-horizon work, and that early results point to stronger performance and better cost efficiency. The firm has also said the model was developed with research support from Fireworks, a software company focused on model deployment and optimisation.
The shift is significant because Harvey has previously relied on models from Anthropic, OpenAI and Google to power legal applications for large law firms and enterprise clients. That earlier approach reflected a broader industry pattern in which start-ups layered specialist tools on top of the biggest proprietary systems. Harvey’s use of Kimi K3 marks a different path: building domain-specific intelligence on an open-weight base that can be further adapted to a narrow professional use case.
Moonshot AI’s release of Kimi K3 weights has added momentum to this trend. Reporting from Tom’s Hardware said the model is positioned as a lower-cost alternative to leading frontier systems, with claims of comparable performance at significantly lower running costs in some workloads. For legal-tech developers, that combination of openness and efficiency is attractive. As AI policy researcher Simon Hedlin noted on X, open-weight models can let companies refine systems on private industry data while reducing inference costs, which is especially relevant in sectors where accuracy and deployment economics matter as much as raw model scale.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
