Story perspectives

Harvey launches Tenet using Moonshot’s Kimi K3 model

8/22/2026

Summary

Story summary

Harvey, a U.S. legal-tech startup, built its first in-house model, Harvey Tenet, using Moonshot AI’s Kimi K3.
Harvey said the model achieved state-of-the-art performance on complex legal tasks.
AI policy researcher Simon Hedlin called the pivot an example of open-weight models enabling higher accuracy and lower inference costs.
The shift reflects a broader move by Western firms toward Chinese open-weight AI amid rising development costs.

History

Historical Context

Google’s release of the encoder-only BERT model in October 2018, with 340 million parameters and publicly released weights, sparked the shift toward open-weight transformers. That shift accelerated after 2023 as open-weight LLMs like China’s DeepSeek-R1 (671 billion parameters, 2025) gained traction. Today Harvey’s legal-tech startup adopts Moonshot AI’s Kimi K3, a Chinese open-weight model, echoing the 2018 pivot and prompting U.S. firms and regulators to reconsider AI sourcing.

1 of 2

Story summary

Harvey, a U.S. legal-tech startup, built its first in-house model, Harvey Tenet, using Moonshot AI’s Kimi K3.
Harvey said the model achieved state-of-the-art performance on complex legal tasks.
AI policy researcher Simon Hedlin called the pivot an example of open-weight models enabling higher accuracy and lower inference costs.
The shift reflects a broader move by Western firms toward Chinese open-weight AI amid rising development costs.

1 / 2