Harvey, the San Francisco-based legal-AI company used by more than 200,000 lawyers across 2,400 organizations, has introduced Tenet, its first post-trained open-weight model for legal work. The Aug. 20 research preview is built on Kimi K3, the open-weight foundation model released by Beijing-based Moonshot AI（月之暗面）, and was developed with AI-infrastructure company Fireworks.
The project is a notable example of a US vertical-AI company building on a Chinese foundation model rather than solely adapting closed systems from OpenAI, Anthropic or Google. Harvey said it had previously focused on tailoring those proprietary models for legal applications. With Tenet, it is testing whether open-weight models can provide greater control over training, deployment and inference costs in high-stakes professional workflows.
Harvey post-trained Kimi K3 for long-horizon legal tasks using a combination of synthetic data, publicly available legal data and human expert data. The company said its training programme used about 150 Nvidia B300 GPUs over two months and relied on approximately 1,750 simulated legal-work environments. These tasks were designed around activities such as contract review, M&A diligence and litigation-related analysis, with model outputs assessed against attorney-developed rubrics.
Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model, with 104 billion parameters activated per token and a one-million-token context window. Moonshot released the model’s weights in July, enabling outside developers to download, modify and deploy it. That openness is central to Harvey’s experiment: it allows the company to post-train the underlying model rather than merely prompt or fine-tune a closed provider’s API.
Harvey said Tenet completed nearly twice as many held-out tasks as base Kimi K3 on its Legal Agent Benchmark and scored strongly on several legal benchmarks. However, the company also notes that much of the evaluation was conducted in its own environment, and results can change with the model harness, grading method and tool configuration. Harvey says it did not use customer data in its post-training work.
The announcement comes as Harvey scales rapidly. The company says more than 75 Am Law 100 firms use its platform, while a March funding round valued it at US$11 billion. The Information has reported that Harvey is now discussing a new round at roughly US$15.5 billion, with annualized revenue above US$350 million.
For Moonshot, Tenet provides a meaningful overseas reference case: Kimi K3 is not simply being offered to foreign users through an API, but is serving as the base for specialised model development at a major US legal-technology company. K3 has also drawn attention in independent and semi-independent evaluations, including strong results in Arena’s blind front-end coding tests and a sixth-place ranking in a Nikkei–Weights & Biases survey.
The broader implication is less that Chinese open-weight models have become the default choice for Western enterprises than that they have become plausible candidates for specialised post-training. Tenet remains a research-stage effort, and its legal performance must still be tested in real client matters under professional oversight. Yet it shows how open model weights can turn a foundation model into an exportable technology layer—one that overseas companies can adapt to their own data, workflows and compliance requirements.
Contact Us:
Xingyiran@iyiou.com
