Harvey Tenet: What the First In-House AI Legal Model Means for Builders

Harvey has launched Tenet, its first proprietary model for legal work, post-trained on Moonshot's Kimi K3 open-weight base source. The move shifts Harvey from renting inference on third-party models to owning its engine, a change that directly affects cost structure, vendor dependence, and regulatory risk for the legal tech firms and their law firm customers.

What Tenet Is and How It Was Built

Tenet is a post-trained derivative of Kimi K3, an open-weight model released by Chinese startup Moonshot. Harvey worked with Fireworks AI on the training source. To shape the model for legal work, Harvey's in-house lawyers and external partners like Mercor and Snorkel created mock disputes and case files, then graded the model's reasoning on those fabricated scenarios.

The company's cap table makes the move more pointed. OpenAI is an investor in Harvey alongside Sequoia and Andreessen Horowitz source. Tenet is designed to displace GPT-based workflows within Harvey's own product, converting what was a variable invoice to a competitor into a fixed licensing cost.

What Changes for AI Builders Using Harvey

The practical shift for builders is in two areas: cost exposure and counterparty risk. Every model call Harvey used to make on a customer's behalf was a transaction with a model provider that could change pricing, capabilities, or terms at any time. Owning Tenet turns that into a license negotiation with Moonshot.

The Kimi K3 license permits derivative models, but it requires a separate agreement with Moonshot for model-as-a-service operators whose revenue exceeds $20 million in any twelve months source. Harvey's annualized revenue is above $350 million, so the terms of that upstream agreement matter directly. For builders integrating Harvey into a larger stack, the licensing of the base model becomes a compliance input, not just a technical one.

Regulatory and Cross-Border Implications

For law firms operating in Europe, the EU AI Act adds another layer. The Act's guidelines use roughly a third of the original training compute as the marker for a significant modification that would shift provider obligations downstream source. A post-train like Tenet almost certainly falls well below that threshold, meaning most obligations stay upstream with Moonshot. But the practical effect is that a European firm deploying Tenet must rely on a documentation chain that starts in Beijing. That creates due diligence challenges around data handling, model behavior, and audit trails that may not align with European regulatory expectations.

The choice also raises a question about privilege. Harvey has effectively placed Chinese open weights underneath data that may be covered by legal privilege. For law firms and their clients, the counterparty risk in the model supply chain is now a factor in client engagement letters and risk assessments.

Caveats

The performance claims for Tenet come from Harvey's own research, which showed the model performing competitively against frontier models from Anthropic and others source source. Independent benchmarks have not yet validated these results in legal-domain tasks. Pricing for the model and the exact timeline for replacing GPT workflows within Harvey's product have not been disclosed. Builders should treat the licensing and compliance details as directional until Moonshot and Harvey publish definitive contractual terms.