Harvey has been publishing legal benchmarks for two years. BigLaw Bench, then LAB, the Legal Agent Benchmark. Slow, unglamorous work, mostly read by other people building legal AI.

That is the reason it could build a model last month.

On August 20 Harvey released Tenet, post-trained on Moonshot AI’s open-weight Kimi K3. First place on LAB Contracts. Second on LAB overall.

Second, not first. The company said so itself, and most of the coverage rounded it up into a story about a Chinese model beating the American labs. The honest number is the interesting one, because it tells you what kind of claim this is.

What the report actually says

Harvey’s write-up reports that Tenet completes almost twice as many held-out tasks on LAB as the base Kimi K3, and 20% more on LAB Contracts. All-pass rate climbs 9 percentage points on LAB and 2 on LAB Contracts.

Then the part that carries the weight. Those gains transfer to two benchmarks Harvey does not own and the model never saw in training, Mercor’s APEX Agents for corporate law and Crosby’s Redline Bench.

A model that only improves on its own benchmark has learned the exam. A model that improves on somebody else’s has learned the work. That distinction is the whole difference between a press release and a result.

The post-training ran with Fireworks, as reinforcement learning over roughly 1,750 legal tasks. Each task carries an expert rubric, 50 pass/fail criteria on average, and a judge model grades the work against it. Harvey built the human expert data with Mercor. The expensive part is people who know the law writing down what good work looks like, criterion by criterion.

Owning the evaluation is the precondition

Any software company can download Kimi K3 this afternoon. The weights are public and the tooling is commodity.

Almost none of them can answer the next question, which is whether the tuning worked. Answering it takes a body of tasks from your own industry, graded by people who know the industry, large enough and honest enough to separate a real gain from a model that learned to sound more confident.

Without that, post-training is spending money into a hole.

Harvey had the instrument before it had the ambition. Measure first, train second. Reversing that order is how most enterprise fine-tuning projects end up with a model everyone agrees feels better and nobody can defend in a review.

We came at the same idea from the other side in what cannot be trained, which asks which part of your judgment never made it into the data. This is the follow-on. If you can write that judgment down as an exam, it becomes an asset no vendor gets to price.

The Chinese base matters less than the license

There is a geopolitics conversation around this and I understand why it exists, but for an architecture decision it is mostly noise. The weights are open. Tenet is still a research preview, and when Harvey moves it into production, what serves customers will be Harvey’s tuned copy.

The license is the part worth reading. Moonshot requires a separate agreement for model-as-a-service operators past $20 million of revenue in twelve months. At roughly $350 million annualized, Harvey is far past that threshold, so before Tenet reaches customers it has to settle whether its use falls under that definition.

Check the license before you fall in love with the model. It is boring and it is where half of these projects quietly die.

The underlying pattern showed up in voice first, when Fish Audio undercut the market by 70% and ElevenLabs grew anyway. Cheapening the model relocates the value upward, into the product and into the evaluation data.

If you sell vertical software

Three questions I would put to your team this week, and I mean them literally.

How many real customer tasks do you have graded by someone qualified to grade them? Were the criteria you grade them against written by an expert or by another model? What did you spend last month on API calls for work that repeats almost identically?

If the first answer is zero, an open model is not yet useful to you. The weights are fine. You just have no way to find out whether they helped.

That is where our Technology Partner work with software companies starts, building the evaluation set for your vertical before touching a single model weight. Harvey took two years to get theirs. You do not need two years. You need to start at the same end.

Build the benchmark for your vertical