Harvey trained Tenet, its first in-house legal model, on Moonshot's open-weight Kimi K3 instead of rented US lab systems. Nigeria's Curacel already routes high-volume claims and fraud work to Chinese models.
Harvey released Tenet, the first model it trained in-house for legal work. The system is post-trained on Kimi K3, the open-weight model the Chinese startup Moonshot released in July, rather than on the OpenAI, Anthropic and Google systems the company had been renting. Fireworks AI worked with Harvey on that post-training. Tenet is positioned to sit under client legal work that until now ran through those rented models.[1]
For evaluation material Harvey hired lawyers and worked with Mercor and Snorkel on synthetic legal cases. OpenAI is an investor in Harvey, which carries a valuation of 15.5 billion dollars. No benchmark figures were published alongside the release.[1]
Curacel, a Nigerian firm that sells claims and fraud-detection services to African insurers and fintech companies, sends high-volume coding, data extraction, classification and customer support to Chinese models. Chief executive and co-founder Henry Mascot named GLM-5.3 from Zhipu AI, DeepSeek, Alibaba's Qwen and Moonshot AI's Kimi. He said the leading Chinese models have closed much of the performance gap, while Western systems remain stronger for demanding reasoning and reliability-sensitive work. Curacel keeps more than one model connected and routes each workload by quality and cost; the account of what the models deliver comes from the company itself.[2]
