Why Harvey Built Tenet on China’s Kimi K3 Instead of Renting AI Models

How Harvey trained Tenet

Tenet is not a new general-purpose foundation model trained from the ground up. It is a specialised model created by taking an existing open-weight base—reported as Moonshot AI’s Kimi K3—and applying additional training for legal reasoning.

The reported training process combined:

• publicly available legal material;
• synthetic legal tasks and simulated disputes;
• mock case files and contract scenarios; and
• examples created, reviewed or graded by lawyers and other legal experts.

That last component is important. Legal expertise is not only a collection of documents; it also involves judging whether a proposed interpretation, negotiation strategy or multi-step analysis is sound. Having attorneys create scenarios and assess model outputs gives Harvey training and evaluation data aimed at the quality standards of actual legal work.

What Harvey says about Tenet’s performance

Harvey describes Tenet as reaching “frontier-level” performance on prominent legal benchmarks and performing on par with leading general models for the relevant work.

Its own technical update reports that Tenet completed almost twice as many held-out Legal Agent Benchmark tasks as the Kimi K3 base model, while completing 20% more tasks on LAB Contracts. Harvey also reported increases in all-pass rates of 9 percentage points on LAB and 2 percentage points on LAB Contracts.

A separate company announcement presents the gains as relative improvements of 82% on LAB and 22% on LAB Contracts, and says Tenet ranked first on LAB Contracts and second overall on LAB.

These figures should be read as benchmark claims, not as proof that Tenet is universally more accurate than closed models or reliable for every jurisdiction and legal matter. The supplied evidence is primarily Harvey’s announcement and reporting based on it; independent validation of real-world legal accuracy is not established here.

From model integrator to model owner

Before Tenet, Harvey’s product strategy was primarily application-led. It could customise prompts, workflows and routing while using proprietary models supplied through external providers. That approach can offer access to powerful systems without requiring the company to build and operate its own model stack.

Tenet changes the position of the model inside Harvey’s technology stack:

• Previous approach: orchestrate or customise closed models through external providers.
• Tenet approach: start with released model weights, post-train them on legal tasks, and use the resulting specialised model within Harvey’s product.

The distinction is not simply technical. With access to the weights, Harvey has more freedom to tune the model against its own data, evaluations and product workflows. It can also make decisions about serving infrastructure and optimisation instead of paying a model provider for every request through a closed API.

What post-training means

Post-training is additional training performed after a model’s broad initial training is complete. Instead of teaching general language ability from scratch, developers use curated examples, expert feedback, synthetic tasks and outcome-based evaluations to improve particular capabilities or behaviours.

For Tenet, the target capability was legal reasoning over long-running tasks. Harvey’s reported recipe used synthetic, public and human-expert data to teach and test how the model handles legal scenarios, then measured the results on held-out tasks.

This approach can be more efficient for a vertical-AI company than building a new frontier model. The base model supplies broad language and reasoning capabilities; domain-specific post-training focuses additional effort on the workflows that customers actually use.

Why open weights can improve specialisation and cost control

An open-weight model makes its numerical parameters available for an organisation to run and modify, unlike a closed model that is accessed only through a provider-controlled API. Kimi K3’s released weights gave Harvey a base it could directly adapt for legal work.

That control can support more targeted accuracy. A company can repeatedly train and evaluate against the documents, task sequences and quality criteria that matter to its industry or customers. The result is not automatically better than a closed model, but it gives the company more freedom to optimise for a narrow production use case.

Open weights can also change the economics of inference. A company can select its own hardware, serving software and optimisation techniques rather than paying a closed-model provider’s price for every token. The model is not free to operate: compute, hardware, engineering, security and maintenance still create substantial costs.

For legal agents that may process large document sets or run for extended periods, reducing the cost of each operation can make continuous or multi-step workflows easier to offer commercially. That is the business logic behind combining a capable base model with domain-specific post-training.

What Harvey’s pivot says about the AI market

Harvey’s move is a notable example of a wider change in vertical AI. Application companies do not necessarily need to win by training the largest general-purpose model. They can instead compete through specialised data, expert feedback, evaluations, workflow design, deployment and customer trust.

The choice of Kimi K3 is also strategically significant because it shows that a Western company can treat a Chinese open-weight model as technical infrastructure even while the leading AI labs remain concentrated in a broader geopolitical and commercial competition. The attraction is not proof that Chinese models are universally superior. It is that open availability, customisability and potentially lower serving costs may matter more than a model’s position on general-purpose benchmarks for a focused enterprise product.

The caveat is equally important: open weights transfer control, not all costs or risks. Companies still need the infrastructure to run the model, the expertise to evaluate it, and safeguards for sensitive legal data. Tenet therefore represents less a rejection of closed models than a bet that specialised model ownership is becoming practical—and strategically valuable—for enterprise software companies.