A Legal AI Company Built Its Own Model, and It Started From an Open Chinese One
Harvey launched Harvey II with a Memory feature that learns each lawyer's drafting style, plus Tenet, its first in-house legal model, post-trained from a version of Moonshot's open-weight Kimi K3.
Harvey, the legal AI company that has spent three years selling access to other people’s models, announced on Tuesday that it now has one of its own. Harvey Tenet is the company’s first proprietary model trained specifically for legal work, post-trained on mock disputes and case files starting from a version of Moonshot AI’s Kimi K3. Harvey says Tenet reaches frontier level on prominent legal benchmarks, performing on par with the strongest general models at what it describes as an open-source cost.
Tenet arrived alongside Harvey II, a broad refresh of the platform. The centrepiece there is Memory, which learns an individual lawyer’s drafting style and preferences and carries them across Harvey, Word, Outlook and the company’s agents. There are also Spaces, collaborative areas organised around a matter or project, and agents that start work with the files, context, permissions and history of that matter already loaded rather than needing to be told everything from scratch.
The model choice is the interesting bit. Kimi K3 is an open-weight model from the Chinese lab Moonshot AI, meaning the trained parameters are published for anyone to download and modify. Post-training, in plain terms, is what you do after the expensive part is finished: you take a model that already knows how language works and teach it the habits of a specific domain. Harvey did not spend a fortune building a model from nothing. It took a strong open base and made it a lawyer. That is a route the biggest labs cannot offer, because they do not hand over their weights, and it is now the standard playbook for any company with deep domain data and no appetite for a billion dollar training run.
Two things to keep grounded. “Frontier-level on prominent legal benchmarks” is Harvey’s own claim on Harvey’s own framing, without independent evaluation, and legal benchmarks are a thin and contested field. And Memory features are genuinely useful but also genuinely sticky: a system that has learned how you write is a system you will find harder to leave. That is not a scandal, it is the business model, and it is worth naming.
What this means for you: if you work anywhere near law, expect your tools to stop being generic chatbots wearing a legal skin. The direction is systems that already know the matter you are on and the way you write. More broadly, this is a good illustration of why open-weight models matter even if you never download one. They let smaller companies build serious specialised products, which means more competition and lower prices for everyone downstream. If you run a business with a lot of domain-specific documents, the Harvey route is now technically within reach for far less money than it was a year ago. Whether it is worth it depends entirely on how much of your work is genuinely unlike everyone else’s.
Sources
Microsoft Has Spent 280 Billion Dollars on AI. The Guardian Asked Where the Chips Are.
Internal documents reportedly put Microsoft at 2.2 million AI chips installed, well below what its stated 5GW of data centre capacity implies. Microsoft calls the analysis inaccurate. Nadella blames power.
