Harvey raises $550m at $15.5bn, and its own model runs on Chinese weights

Harvey raised $550m at a $15.5bn valuation on 9 September, in a round co-led by Diffusion and Lightspeed that arrives six months after $200m at $11bn and nine months after an $8bn mark. Total raised is now above $1.55bn. PitchBook counts at least eight priced rounds since 2023, five of them since 2025, and Harvey, like Databricks, declined to give this one a letter (by the count of extensions it is roughly a Series F, which is a good measure of how little the letters now mean).

Do the arithmetic on the step-ups and something odd falls out. In March, $200m of new money arrived attached to $3bn of new valuation — fifteen dollars of markup for every dollar in. In September, $550m bought $4.5bn, or $8.20 per dollar. The company got more valuable and the exchange rate got worse.

That is what a round looks like when the seller is taking capital it does not obviously need.

The interesting sentence is in Harvey's own post, not in the coverage. The company says the raise "follows the introduction of Harvey's first post-trained open-weight model and the launch of Harvey LAB, the Legal Agent Benchmark." That first item is Harvey Tenet, announced a couple of weeks earlier: built from Kimi K3, Moonshot's open-weight model, post-trained on legal data with help from Fireworks (the same inference shop that helped Cursor train its own model, which tells you something about who actually builds these things). Harvey is also telling customers to post-train open-weight models of their own.

So the flagship model of a $15.5bn American legal company sits on a Chinese base. Nobody at the announcement seemed to find this worth a sentence.

We would. Not for the flag — for the licence. Moonshot's published terms for Kimi K3 require a company running a Model-as-a-Service business to reach a separate agreement with Moonshot once it and its affiliates clear $20m of revenue in any twelve consecutive months, and the provision reaches derivative models. Harvey sells legal software, not model access, so it is probably outside that clause. Probably. As far as we can tell nobody has tested where a post-trained derivative sold inside a product ends and a model service begins, and the firm most likely to find out is the one that just raised half a billion dollars to scale exactly that arrangement.

The strongest counter is that the base is interchangeable, so none of this is a dependency: Harvey's moat is workflow, data and 80% of the Am Law 100, and it could swap Tenet's foundation in a quarter. True, and it is why we would not price this as a risk. But interchangeable is not free. A swap costs an evaluation cycle, a re-post-training run and a customer conversation about why the answers moved, which is a bad quarter to have while you are converting the last twenty firms on the list.

Our read is that Harvey has bought independence from OpenAI and Anthropic at the price of a dependency nobody has priced yet, and that the second one is cheaper. The party worse off here is the frontier lab enterprise sales team: law was supposed to be the showcase vertical for premium API access, and the showcase account just built its own model on someone else's weights and told its customers to do the same.

What would move us off that: Harvey's next in-house model going back to a closed frontier base, which would say the open-weight route did not hold up under legal evaluation.

One number to keep. Harvey says 80% of Am Law 100 firms use it, along with five of the Fortune 10 in-house teams. At $15.5bn, the market is pricing something well past the Am Law 100 — there are only twenty of those firms left to sell. Which market is that, and does Harvey have to leave law to reach it?