Skip to content

Blog/Trust & product

You no longer have to send the file away to get a strong model

24 October 2026 · 6 min read

For two years, a firm that wanted a strong model had one practical choice. Send the work to the company that owns the model. The open models, the ones a firm can run for itself, were not good enough for professional work. We know that, because we started building before they were.

In 2026 that excuse is gone. On a published legal test, several open models now sit close to the models firms have been paying to call over the internet.

What the published numbers say

Vals AI keeps a Legal Research Bench. The page we read was updated on 27 October 2026. It asks whether a model can do legal research across United States federal and state law. It is not a test of Hong Kong law, and it is not a prospectus, a facility agreement or a hearing bundle.

On that date the highest score was 55.77 per cent. Just over half. GLM 5.3, from Zhipu, scored 49.04 per cent. Qwen 3.8 Max, from Alibaba, scored 47.60 per cent. A recent DeepSeek model, V4.1 Flash, scored 45.19 per cent. MiniMax sits further back on this particular test, at 29.81 per cent.

They are not interchangeable, and none of them replaces the lawyer. Even the leader was wrong almost half the time. The useful fact is narrower. On a hard, published legal-research test, the better open models are now in the same range as the famous closed ones.

What that changes

The old commercial argument was that nothing else could do the work, so the file had to be sent. Latham & Watkins has already acted on the new position. It is adapting Nvidia’s open models on servers the firm controls, for client information it does not want to give to a cloud vendor.

“We have to use a closed model” is no longer a technical fact. It is a choice about where the work runs, and about who else can see it. A firm can have a capable engine and still keep the file.

A model is not the system

We would not tell a partner that an open model is as good as a lawyer, or that a leaderboard is a reason to stop reviewing. A score on American research says nothing about a Hong Kong ordinance, a listing rule, or the last mark-up this client accepted.

A model on its own also does not know the firm. It has not read your precedents, your playbook, or the way a particular team writes. That knowledge is what makes a draft usable, and it is also what you should not upload to a model every other firm can call.

Makebell is the system around the engine. The model runs on the firm’s machines. The firm’s own know-how is connected to it, and stays there. Practice material prepared by lawyers — legislation, cases, regulatory documents, and the materials a practice actually uses — is there as source material, so the firm is not starting from an empty chat box. The lawyer still checks the result. The difference is that getting a strong model no longer means handing the file to someone else.

Sources

  1. Vals AI, Legal Research Bench — Updated 27 October 2026. US legal research. Top score 55.77%. GLM 5.3 49.04%. Qwen 3.8 Max 47.60%. DeepSeek V4.1 Flash 45.19%. MiniMax-M3 29.81%.
  2. Financial Times, 10 September 2026 — Latham is fine-tuning open models on its own Nvidia servers so some client information need not go to a cloud vendor.