Trevor Slette

Co-founder, Quadd.ai · Reply to me at trevors@quadd.ai

First, the heads-up: you'll own your AI sooner than you think

Here's the development worth your attention this week, underneath the geopolitics: a Chinese lab called Zhipu released GLM-5.2, a free, open AI that performs within about a point of Claude and ChatGPT on demanding work — with a no-strings license that lets anyone download and use it. For the first time, a genuinely capable AI isn't only something you rent from a big company by the month. It's something that, on the right hardware, you could own outright and run privately. We're not all the way there yet — but my honest bet is that by the end of this year, a model good enough for real newsroom work will run on hardware a small paper can actually afford.

Owned, local AI means your most confidential documents could be analyzed by a capable model without a single byte ever leaving your building. For source protection, that's the holy grail — and it just moved from "fantasy" to "matter of months."

How good is it, really? Surprisingly close to the best

The fair question about any free AI is whether it's actually good, or whether you get what you pay for. The honest answer here: GLM-5.2 is startlingly close to the best money can buy. Across a wide range of real coding and reasoning tasks it lands within about a point of Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 — close enough that on everyday work you genuinely couldn't tell which one wrote the answer. It beats Google's Gemini on coding outright, and it's the top-ranked open model on essentially every test out there. The paid models only pull clearly ahead on the very hardest, longest jobs — and GLM-5.2 costs roughly a sixth of what they charge.

Translation for a newsroom: "free and open" is no longer the compromise choice. For summarizing documents, drafting, and analysis, you're choosing between near-equals — and the open one is the only one you can eventually run on your own machine.

Now the honest part: the flagship costs about $300,000 to run yourself

I'm not going to hype this. Running the actual GLM-5.2 in your newsroom isn't realistic today, and anyone who tells you otherwise doesn't know the numbers. It's a giant — running it takes roughly eight specialized AI chips wired together, a setup that costs about $300,000 to buy. You could rent that horsepower from a cloud provider instead, but that's still around $15,000 a month running full-time — and to make even renting cheaper than Zhipu's own $30-a-month hosted plan, you'd have to be firing thousands of requests a day. Even if this hardware gets ten times cheaper over the next few years, owning the full flagship still lands near $30,000. So: a heads-up, not a to-do — but the curve is bending fast, and the smaller open models are the on-ramp you can step onto right now.

Be skeptical of anyone urging you to self-host the flagship today — the math doesn't work yet for a community paper. But the trajectory is real, and getting familiar with the smaller, runnable models now is how you'll be ready the day the price meets the moment.

How to run an open model yourself (the exact steps)

Here's the part to actually try — and you're not limited to the Chinese model in the headlines. There's a whole shelf of open models you can download and run free on your own computer: Google's Gemma (US), Meta's Llama (US), and Mistral (France), right alongside GLM and Qwen. Which one you pick depends mostly on how much memory your computer has, and the whole setup takes about 20 minutes with no coding. The full walkthrough covers which free app to install, how to choose a model that fits your machine, how to run it completely offline, and which options to use if you'd rather not run a Chinese model at all.

The point isn't to deploy a giant flagship today. It's to feel what owned, private AI is actually like on your own machine — so when the hardware catches up to the price, you already know exactly what you want and how to run it.

See the exact step-by-step →

— Trevor

Trevor Slette

Co-founder, Quadd.ai · 28-year community publisher

trevors@quadd.ai · Book a 15-min call · LinkedIn

P.S. P.S. Want me to flag you the day a model as capable as GLM-5.2 can run on hardware a community newsroom can actually afford? Reply "local" and I'll keep you on the watch list — that's the moment everything changes for source protection.

Run a newspaper. Use Quadd.

AI tools built for small newsrooms. Document intelligence that turns court reports and box scores into copy you can paste into your layout. Audio transcription that finds the quote without you scrubbing for it. AP-style proofreading that tightens copy without flattening voice. Built by someone who knows what a Tuesday night feels like. Free for seven days.