Local AI finally went mainstream — and your password generator should too
I did something stupid last week
Last week I was spinning up a local model and needed an API key for an external service. I opened the first password generator that search surfaced, ticked a few boxes, hit generate, copied the string, moved on. Five minutes later it hit me: I had just shipped a secret that should never leave my machine to a server I had never heard of. The irony stung. In 2026 we scramble to run AI on phones and laptops precisely so data stays put, then we hand the most sensitive string we own right back to the cloud.
24 billion parameters, and the pitch is staying put
The local-first shift is not hype. On October 14 Liquid AI shipped LFM2.5-24B-A2B: 24 billion parameters, but only 2 billion active per step through a sparse MoE design, decoding at 184 tokens per second on an M5 Max with a 131,072-token context window, weights open. The sell is offline agents on hardware you already own. The same day FLock.io released THIS/THAT, built on Qwen3.5-2B at 1.88 billion parameters, 31.4-millisecond local latency, 87.8 percent on 1,710 decision tasks, 2.5 million inferences in three weeks. NVIDIA's Underdog app for RTX Spark devices followed, claiming roughly 1.4x faster end-to-end agent runs. The throughline is identical: keep the data on your own metal.
Passwords should be the last local line
That keep-it-local framing is exactly why these models sell. Yet the way most of us generate passwords, tokens, and API keys violates the same principle daily. You pick rules on a web page, click generate, and that high-entropy string travels over an HTTP request alongside your preferences. Whether the server logged it, profiled it, or got breached later is information you never get. Plenty of those pages also drop a tracking script while they are at it.
Antpact's password generator takes the other path. The rules are set in your browser, the randomness is computed locally, and nothing you type or receive is sent back to a server. Length, character sets, ambiguous-character exclusion such as 0 versus O and 1 versus l, pick and click, then close the tab with nothing left behind. The same local-first design runs through the rest of the toolbox: the UUID generator, the TOTP authenticator, the Base64 tool, all compute in the browser and discard on close. The tool lives at antpact.com/text/password-generator/.
Two habits after the mistake
My routine is now fixed. When a local model needs an external key, I generate it in the browser tool at 24-plus characters with all four character classes on, paste straight into the config or a secrets manager, and never route it through anything online. If I need to verify a hash, I use the local hashing tool rather than dragging a file into a web form. You do not upload a 12GB model weight just to compute its SHA256.
Two rules came out of that mistake. Anything that asks you to paste content to encrypt or text to sign is collecting samples, so do it locally when you can. And do not screenshot a generated password into a group chat, because screenshots leak far more often than the plaintext ever would. Neither is exotic. It is just refusing to hand the most private data to the least appropriate party.
Does that string actually need to be told to anyone
The on-device AI wave turned data stays on the device into a default expectation. Passwords and keys were always the category that most deserved to stay local. There is no reason to break that chain at the last step. Next time you need a strong password, ask the obvious question: does this string actually need to be told to anyone else?