Make your own AI model that follows your rules
Write down how it should answer, like the format to use or things it must never say, and Weightship trains a small model that keeps those rules without being reminded. It keeps them even when someone adds their own instructions on top, which is where a prompt usually gives way.
It suits narrow jobs you need done the same way thousands of times: support replies, answers in a fixed format like JSON, sorting and labelling. For open-ended reasoning a large general model is still the better choice. We host your model and you pay per use, and if you like, other people can use it too and you earn from each request.
How it works
Make a model
Start from a list of rules, like "always reply in JSON" or "never mention competitors", or from 50 to 500 example answers. We train it on our GPUs and you pay for the minutes it takes. If you already have a fine-tune on Hugging Face, paste the link instead.
Use it yourself
Call it from your app with the code you'd write for OpenAI, changing only the address and the model name. The model is yours, so you can download its files whenever you like.
Sell it
List it on the marketplace at a price you choose. Customers pay for what they use, we take out hosting and a 20% fee, and the rest is paid to you every month through Stripe.
Ask both models the same question
We trained a small open model, Qwen3-4B, on the rules below and then stopped giving it any instructions. On the left is the same model before training. Pick a question to see how each one answers, and tick the box underneath to give the untrained one the rules as well.
- Start with the answer, no "Sure!" or "Great question"
- Plain text, no headings, bullets or bold
- Under 120 words
- No em dashes
- Don't end with "Let me know if..."
We asked both versions 120 questions they hadn't seen in training. With the rules pasted into its instructions, the untrained model kept all five writing rules on 88.3% of answers and the trained one on 97.5%. Once a customer added instructions of their own, that became 65% against 92.5%. How we tested it
Who it's for
A business
You want a support assistant that always answers in your format and never promises refunds. Train those rules in once, and they still hold when other instructions get added on top.
Someone with a niche
You know a subject well or have examples nobody else has. Turn that into a model and charge people to use it.
A researcher
Compare training methods on your own tests and publish the results as a page anyone can rerun.
What it costs
Signing up is free and there's no monthly fee. You add credit from $5 and it's used as you go.
For a small model, paid by the GPU minute. Before a run starts you see the most it can cost, and it stops rather than go over.
Per million tokens, which is roughly 750,000 words read and written. Calls draw from your credit, and you can cap what each API key spends in a month.
Tests against other providers' models run on your OpenAI, Anthropic or OpenRouter key, so they bill you directly at their prices.
Earn when other people use your model
Put your model on the marketplace and set a price per token. Hosting is the GPU time your model actually used, shared with the other models on the same machine, so a model that gets little traffic costs little to keep online.
Common questions
Do I need my own GPU?
No. Training and hosting run on our GPUs. You pay for the minutes a training run takes and per token when the model answers. If other people use it, its hosting comes out of what it earns.
Who owns the model?
You do. You can download the trained files at any time and run them anywhere the base model's licence allows.
What if nobody uses it?
Then it costs you nothing after training. A model with no traffic uses no GPU time, so there's nothing to take out.
Why not just put the rules in the prompt?
A prompt has to travel with every request, and anything a customer adds to it can push your rules aside. Training puts the rules into the model itself, which is the difference the comparison above shows.
How does the training work?
The base model answers a few hundred questions with your rules in front of it. We keep only the answers that pass your checks and teach the model from those, so afterwards it keeps the rules with nothing in front of it. The blog has the full write-up and the numbers.