Legal AI startup Harvey has shifted its strategy toward open-weight models after experiencing a sharp decline in gross margins, which fell from 50% to negative 50% within six months following a product update in March 2026. The surge in customer usage led to unsustainable inference costs from proprietary AI providers.
On August 20, 2026, Harvey launched Tenet, its first internally post-trained model based on the open-weight Kimi K3 architecture. Kimi K3, released in July 2026, features approximately 2.8 trillion parameters. By controlling its own infrastructure for model deployment, Harvey aims to replace per-token API fees with compute-based spending.
The company continues to grow despite its margin challenges. By September 2026, Harvey reached a valuation of $15,5 billion, up from $11 billion in March 2026. It also closed a funding round of $550 million and currently reports annual recurring revenue exceeding $400 million, serving over 3,000 customers.
Harvey’s platform now utilizes specialized benchmarks, such as BigLaw Bench, to route tasks between models based on performance and cost. This transition marks a strategic move from competing solely on application-layer quality to developing proprietary models, reducing reliance on external providers like OpenAI and Anthropic.
