On 13 June a Chinese lab released GLM-5.2 — an open-weight model that scores just behind the best closed model in the world on serious coding benchmarks, carries a million-token memory, and comes with a licence that lets anyone download the weights and run them on their own machines. For a business in finance, law, accountancy or healthcare, the headline isn't the benchmark. It's that near-frontier capability no longer has to be reached by sending your sensitive data to someone else's cloud. That removes the single objection I hear most from compliance-bound owners. It does not remove the work of running it responsibly. Here's what just got possible, and what it costs.
If you run a business where client data is the whole liability — a brokerage, a law firm, an accountancy practice, a clinic — you have probably said some version of this sentence to a consultant: “I'd love to use AI for that, but I can't put client information into an American chatbot.” It's not paranoia. It's the correct instinct, and for two years it was also a genuine ceiling. The good models lived in someone else's cloud, and using them well meant your data took a trip there and back. For regulated work, that trip is the problem.
That ceiling moved on 13 June. A lab called Zhipu released GLM-5.2, and three numbers tell the story. It scores 62.1 on a hard real-world coding benchmark and 81.0 on another. That's close enough to the best closed model on the market to be in the same conversation, which open models simply weren't a year ago. It can hold a million tokens of context, roughly five times its predecessor, which in plain terms means it can read a very large pile of documents at once. And it ships under an MIT licence: the weights are free to download, and you're allowed to run them on your own hardware, inside your own building, without sending a single byte to anyone.
That third point is the one that matters for regulated industries, and it's the one the benchmark coverage tends to bury. Five of the ten most capable models in the world are now open-weight. The era where “good AI” and “AI you have to rent from a third party” were the same thing is ending.
Why “on your own infrastructure” is the whole game in regulated work
For most businesses, where the model runs is a technical detail. For a compliance-bound one, it's the entire decision. When you call a hosted AI service, your data is processed on infrastructure you don't control, by a company you have to trust, under terms that can change. For a marketing agency that's fine. For a firm holding client money, medical records, or privileged legal documents, it ranges from awkward to forbidden — data-residency rules, GDPR, sector regulators, and your own professional-indemnity insurer all have opinions.
An open-weight model you host yourself collapses that problem. The data never leaves. There's no third-party processor to add to your privacy notice, no cross-border transfer to justify, no vendor whose breach becomes your breach. You're not asking a regulator to trust someone else's cloud; you're keeping the work in a room you already control. That's not a marginal improvement in the compliance story. It's a different story entirely.
What this realistically unlocks for a smaller firm
I want to be concrete, because “AI for compliance” is the kind of phrase that means nothing. The jobs that suddenly become defensible are the document-heavy, judgement-light ones that every regulated practice drowns in:
- Summarising long, sensitive documents — a 200-page case file, a year of statements, a stack of policy documents — without any of it leaving your network. The million-token memory is what makes this practical rather than a fiddle of cutting things into pieces.
- First-pass review — flagging the clauses, transactions, or entries that a human needs to look at, so your qualified people spend their time on the 10% that's interesting rather than the 90% that's routine.
- Structuring messy inputs — turning scanned forms, emails, and notes into clean records, on your own machines, where the raw material is exactly the stuff you can't post to a cloud.
None of these is the AI making a regulated decision. Each is the AI doing the reading and the sorting so a human can make the decision faster. That's the only shape of this I'd put anywhere near a regulated workflow, and conveniently it's also the shape that delivers the most boring, reliable value.
The caveats, because there always are some
First, provenance. GLM-5.2 comes from a Chinese lab, and for some clients and some regulators the origin of the weights will itself be a question worth asking — not because open weights phone home (they don't, once they're on your hardware), but because procurement and due-diligence policies increasingly care where software comes from. It's a conversation to have deliberately, not to wave away. The good news is that the open-weight field is broad now; if one origin doesn't fit your governance, another option likely does.
Second, this is not a free lunch dressed as one. Running a frontier-class model in-house is a real undertaking — hardware, maintenance, security. For a small firm, the right first step is rarely “build a private AI cluster.” It's usually a small, scoped pilot on one painful document workflow, sized honestly, to find out whether the value justifies the operational weight before you commit to it. Sometimes it does and the privacy gain is transformative. Sometimes a careful hosted arrangement with the right contractual protections is the more proportionate answer. The point is that you now have a real choice where eighteen months ago you had none.
What I'd do this month if I ran a regulated practice
I wouldn't rush to install anything. I'd do something cheaper and more useful: I'd write down the one document-heavy task my team hates most — the file review, the statement reconciliation, the policy summarisation — and I'd ask a simple question of it. If the data never had to leave the building, would I let AI touch this? For a lot of firms the honest answer flips from “no” to “yes, carefully” the moment the cloud trip disappears. That flip is the news. Everything after it is implementation, and implementation is the easy part to get help with.
The capability arriving on a server you control rather than a service you rent is, quietly, one of the more important shifts of the year for anyone whose business is built on confidentiality. It doesn't make the compliance work go away. It just means the answer to “can we use this at all?” is no longer a flat no.