Custom training
Custom training that hands you the weights
Supervised fine-tuning, alignment, continued pre-training and vision tuning, on dedicated GPUs metered by the second. The checkpoint is a file, the file is yours, and the meter stops with the job.
Illustrative — every stage runs on the same platform.
The models worth building on
Reach production
What makes a custom model shippable?
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 a training run looks like
01
Bring the data
JSONL examples for instruction tuning and alignment, or raw documents for continued pre-training. The format is simple on purpose: whatever your tools can export, you can train on.
02
Choose the method that fits
Begin with the cheap one — a LoRA SFT run settles most quality questions for a slice of a full fine-tune’s cost — then move up to full weights or an alignment pass when the evals say the cheap one has plateaued.
03
Serve it here, or take it away
One click puts it on a dedicated endpoint behind the same OpenAI-compatible API; one export takes the checkpoint to any host. Owning the file keeps every later decision yours.
What teams ask before the first run
Who owns the model once training finishes?+
You do. A finished run produces a checkpoint — a file. Download it, put it on a dedicated endpoint behind the same OpenAI-compatible API, or carry it to another host entirely. The weights are the deliverable.
Which training methods are supported?+
Supervised fine-tuning, five alignment algorithms (DPO, SimPO, ORPO, CPO, KTO), continued pre-training, and vision-language tuning — delivered as LoRA or QLoRA adapters, or as a full fine-tune of every weight.
How does training billing work?+
By the second of GPU time on dedicated hardware — no token arithmetic, no seat licences. Backup GPUs can be nominated so a busy first choice never stalls the run. A dedicated GPU meteres while idle, so switch off what you are not using.
How much data do I need?+
Less than you would guess for instruction tuning and alignment — a few thousand clean examples outperform a million noisy ones. For a raw domain corpus, continued pre-training is the method, priced the same per-second way.
Where does my training data go?+
Onto dedicated hardware, never shared capacity, and the resulting weights belong to you alone. Your dataset feeds your job — not anyone else’s model.
Get 50% extra on your first wallet top-up.
Per-second GPU billing · Exported checkpoints · No lock-in