Custom training
Custom training that ends in weights you own.
Supervised fine-tuning, alignment, continued pre-training, and vision tuning on dedicated GPUs billed per second. The checkpoint is a file, the file is yours, and the invoice stops when the job does.
Illustrative — every stage runs on the same platform.
Build with leading models
Reach production
What makes a custom model actually ship?
Three things, and none of them is a demo: ownership you can prove, a method for every quality bar, and billing that respects your budget.
What does a training run look like?
01
Bring your data
JSONL examples for instruction tuning and alignment, or raw documents for continued pre-training. The format is simple on purpose: if you can export it from your tools, you can train on it.
02
Pick the right method
Start with the cheap method — a LoRA SFT run answers most quality questions for a fraction of a full fine-tune — then escalate to full weights or an alignment pass when the evals say the cheap one plateaued.
03
Serve it or take it
One click to a dedicated endpoint behind the same OpenAI-compatible API, or export the checkpoint and serve it anywhere. Owning the file means every later decision is yours.
Questions teams ask before the first run
Who owns the model after training?+
You do. A finished run produces a checkpoint — a file. Download it, serve it on a dedicated endpoint behind the same OpenAI-compatible API, or take it to another host entirely. The weights are the deliverable.
What training methods are supported?+
Supervised fine-tuning, five alignment algorithms (DPO, SimPO, ORPO, CPO, KTO), continued pre-training, and vision-language tuning — as LoRA or QLoRA adapters, or as a full fine-tune of every weight.
How is training billed?+
Per second of GPU time on dedicated hardware — no token math, no seat licenses. You can nominate backup GPUs so a busy first choice does not stall the run. A dedicated GPU bills while it is idle, so stop what you are not using.
How much data do I need?+
Less than you think for instruction tuning and alignment — a few thousand good examples beat a million noisy ones. Continued pre-training is the method for a raw domain corpus, and it is priced the same per-second way.
Where does my training data go?+
Training runs on dedicated hardware, not shared capacity, and the resulting weights belong to you alone. Your dataset is an input to your job, not to anyone else’s model.
Your first run is on us.
New accounts start with a $10 free credit — enough to see the whole loop: data in, checkpoint out, endpoint live.
Per-second GPU billing · Exported checkpoints · No lock-in