Custom models

Not yet available

Self-serve custom models are on the roadmap. If you need to serve your own weights today, contact us — we can discuss it as a dedicated arrangement.

Custom models will let you serve your own weights on our infrastructure, and call them through the same OpenAI-compatible API and keys as catalog models.

How we expect it to work

1. Provide the weights — from a model registry, or uploaded directly.
2. Choose a serving configuration — GPU type, count and precision. Larger models need more VRAM, and quantisation trades quality for cost.
3. We start a dedicated endpoint for it.
4. Call it by its model ID, exactly like any other model.

What we expect to hold true

• Your weights stay private to your workspace.
• Billed on GPU time, not tokens — a dedicated endpoint costs the same whether it is saturated or idle.
• You are responsible for holding a licence that permits your intended use.

This suits models you have fine-tuned elsewhere, architectures we don’t carry, or weights you cannot share with a third-party API.