Every AI model a firm reaches for raises the same question before any other, where does the client's data go and who can read it. A new release sharpened that question on 16 July. Kimi K3, built by the Chinese developer Moonshot AI, entered the independent Artificial Analysis ranking in third place, behind Claude Fable 5 and OpenAI's GPT-5.6, and its weights are due to be published within weeks. The capability is not the point for a law firm. What matters is that a strong, cheaper model from a new jurisdiction will now be offered to you, and your duty of confidentiality meets it unchanged.
What arrived
Kimi K3 is a large model with a context window of about a million tokens, which means it can take in a long bundle or a full contract set in one pass. Moonshot has said it will publish the model's weights by the end of the month, and that lets a technically capable organisation download the model and run it on its own hardware rather than send anything back to the developer. Front-rank quality paired with open weights is what makes this release worth a firm's attention rather than a passing headline. The model also costs less to run than the leading American systems, and that price gap is part of why firms across the market will be offered tools built on top of it. Releases of this size now arrive every few weeks, so a firm meets the same decision again and again rather than once, which is reason enough to settle a way of handling it.
Why the source of a model matters
A model is not a neutral utility. Where it runs, who operates it and which law governs that operator all bear on your duty of confidentiality. When you send a prompt to a hosted model, you disclose whatever the prompt contains to the company running it, wherever in the world that company sits. If that provider falls under a legal regime that can compel access to data held on its systems, your client's information sits within that reach. This is the same test you would apply to any overseas supplier, and it does not soften because the supplier is fast or fashionable. The Information Commissioner's Office sets out what UK data protection law asks before personal data leaves the country, and that reading comes before any trial of the tool.
The open-weight route
Open weights change the picture in a way that helps a cautious firm. A model you can run inside your own environment, or one a supplier runs for you on UK infrastructure, need not send client data anywhere. The prompt stays on systems you can point to and account for. That option carries its own cost and demands real technical care, so it suits a firm working with a capable supplier rather than one installing software itself. The wider point holds regardless. A capable model and strict confidentiality sit together comfortably, once you separate the model from the question of whose servers it runs on.
What a smaller firm should do
You do not need a view on every model that launches, and you should not reach for this one because it ranks well. Treat its arrival as the prompt to check how you handle any tool that touches client information. Before a new model goes near a live matter, know where it runs, who operates it, the law that binds that operator, and what becomes of what you send. Where a tool routes data outside the UK, hold it to the transfer rules rather than to the marketing. A firm that keeps this discipline can take up a strong new model when it fits the work, and turn it down without regret when it does not.
If a new tool has caught your eye and you want to know whether it fits your duties before you let it touch a client file, that is the work our confidentiality and tools review does: see how it fits your firm.