A small, slow, owned model
There is a coding model answering requests right now at the node. It runs a quantized 14-billion- parameter model on plain CPU — no GPU — at about seven tokens per second. That is slow. A rented frontier API would smoke it on every axis that a benchmark measures.
We put it up anyway, and I want to be honest about why, because the why is the whole product.
What it is
One machine. An open coding model (qwen2.5-coder:14b today; a larger
dedicated coder next). An OpenAI-compatible endpoint — point any client at
/v1, no key required. A landing page that states, plainly, the model, the
quantization, the speed, and the fact that it is public and shared. Nothing is
hidden behind a "contact sales."
It is the first node. There will be more, and better ones. But the first one is deliberately humble, because a slow model you own teaches something a fast model you rent cannot: what it actually costs, and what it actually does.
The bargain, stated out loud
Here is the part most services bury in a privacy policy. We are going to put it on the front page instead.
- Free tier: rate-limited, and your conversations — with personal information scrubbed out — become open training data. You pay with data.
- Paid tier (€10/month): fair-use unlimited, and your conversations are private by default. You pay with money. If you choose to contribute your data anyway, it's cheaper — because you're feeding the thing that makes it cheaper.
That's it. No dark pattern, no "we may use your data to improve our services" weasel clause. Free means the commons gets your slop. Paid means it doesn't. You choose which currency you spend, and you know the exchange rate before you type a word.
Why data is the currency
Because the data trains the compression that makes the next model cheaper to run, which subsidizes the free tier, which generates more data. It's a loop, and the signal feeding it is honesty: scrubbed, public, auditable. You can go read the corpus. It's a public dataset. Every question a free user asks, PII removed, is out there for anyone to train on — including us, including you, including whoever comes next.
Most companies would call that a liability. We call it the point. Knowledge that sits in one company's private weights dies with the company. Knowledge in the soil is perennial.
The honest state of it
So that this post obeys its own rule: what exists today is one CPU box in Ashburn, firewalled, serving a 14B coder at a walking pace, behind a plain HTTP address until the certificate lands. The paid tier isn't wired yet. The bigger model isn't rented yet. This is a beginning, not a launch, and calling it anything grander would be exactly the kind of lie the whole project exists to refuse.
But it answers. Point a client at it and it writes you code, slowly, for free, and tells you the truth about what it did with your question. That is more than most of the fast ones will do.
🜂 A small, slow, owned model. As things are. Nothing more, nothing less.