Add my own model

Billed per hour, on a GPU of its own

Your adapter runs on a GPU rented for you alone, with the base model it was trained on. It never shares a GPU with another customer or with the models we serve to everyone else. The price is the live rent of that card plus a margin.

• Adding an adapter is free. You pay only for the hours its card runs, prepaid from your Synap credits. Nothing is rented until the prepayment is debited.
• Your hours start when the adapter answers, not when the card is rented. Starting takes a few minutes while the base model loads.
• No charge per token while it runs. When the paid hours end the card is released; add hours before then to keep it, and unused whole hours are refunded if you stop it early.

Billed per hour, on its own GPU

A full model needs a dedicated GPU, so it is priced by the hour: the live rent of the card you pick, plus a margin always-on or a margin scale-to-zero. Per-token pricing would leave an idle GPU unpaid.

• Always-on runs for the hours you prepay. Auto-renew from your credit balance is optional.
• Scale-to-zero bills in short blocks, only while requests arrive. After a few idle minutes the model is unloaded and the GPU released; the next request starts it again and waits for the cold start (measured on first deployment).
• Prepaid from your Synap credits. Nothing starts until the prepayment is debited. Unused whole hours are refunded when you delete, and you are warned before paid time runs out.

What we can't take

• Pickle weights (.bin, .pt) — safetensors only.
• Models that need custom code (trust_remote_code).
• Architectures we can't serve. Analyze tells you before you pay.
• Private repos. Full models must be public on Hugging Face.