Plan my model What do you need the model to do? A few questions. Your plan updates as you answer: how to get there, which model fits, what to prepare, and what it costs, whether you train here, train yourself, bring a model, or only want to compare.
1 How do you want to work?Do it all here Train on our GPUs, host it, sell it per token Train it myself Get a training kit for your own GPUs or a rented box I already have a model An adapter, a full fine-tune, or a big model Just evaluate Test base models and GPT, Claude or OpenRouter on your rules Run it on my servers Train it here or with the kit, deploy the weights yourself
I'm A business An enthusiast A researcher
2 What should it do?
Smallest that works Best quality, bigger model
In a sentence Rules, one per line Requests your users send (optional, one per line) 3 What do you have?Nothing yet, just the rules A handful of examples Hundreds of examples Thousands of examples
Your plan Base model Qwen3 4B bf16 · shared GPU pool Qwen3 4B is the smallest model that does this kind of task well; it runs on shared GPUs, so you only pay per token served.
run 4-bit3.5 GB
run full9.2 GB
train QLoRA6.8 GB
train LoRA14.8 GB
On our GPUs: we size the hardware for you.
Method LoRA trained from your rules The base model answers a few hundred requests with your rules in front of it, answers that break a rule are dropped, and the rest are distilled in, pushed past what the prompt alone does. On Qwen3-4B that kept five style rules on 97.5% of unseen requests with no prompt, against 88% for the prompt, and 92.5% against 65% under a customer's own system prompt.
What to prepare 5–10 rules, each one sentence (what it must do, what it must never do). 3–10 example requests your users actually send; about 160 practice requests of other shapes are added. Steps We write a project from your rules and a benchmark from your example requests. Train on our GPU (about 15 minutes for a 4B model) and compare it with the base model on the benchmark. Set a price and publish; customers call it through the API. Training about $0.25–$2 of GPU time for a 4B model, more for bigger ones; you pay actual minutes
Serving you set a price per million tokens; the shared GPU cost comes out of revenue, never more than revenue
Add at least one rule you can check, so you can see the model improve.
Create a free account to start Your answers stay in this browser until you act on them.