Deployments
Run models on dedicated GPUs, provisioned and managed by Reka.
Note On-demand deployments are set up with the Reka team rather than self-serve. If you have a workload that has outgrown serverless, or a model we don’t host yet, get in touch and we’ll scope a deployment with you.
Why on-demand
On-demand deployments give you dedicated GPUs for your models, providing several advantages over serverless:
- Better performance — Lower latency, higher throughput, and predictable performance that isn’t affected by other users’ traffic.
- No hard rate limits — Throughput is bounded only by your deployment’s capacity, not by shared serverless limits.
- Cost-effective at scale — Cheaper under sustained, high utilization. Serverless models are billed per token; on-demand deployments are billed for dedicated GPU time.
- Broader model selection — Run models that aren’t available on serverless, and pin a model version for as long as you need it.
- Custom models — Serve your own fine-tuned or custom-weight models for supported architectures, from Hugging Face or your own storage.
How it works
- Tell us about the workloadWhich model (or custom weights), expected traffic, latency targets, and region. A rough shape is enough to start.
- We size and provision itThe Reka team picks the hardware and replica count, deploys the model on dedicated GPUs, and confirms the endpoint and pricing with you.
- You call it like any other modelDeployments are served through the same OpenAI-compatible API, so you point existing code at the deployment and keep going.
When it’s a good fit
Serverless is the right default for most teams: pay per token, no infrastructure to manage, and a catalog that keeps up with new releases. Consider an on-demand deployment when one of these is true:
- You have steady, high-volume traffic where dedicated GPU time is cheaper than per-token billing.
- You need predictable latency or a throughput ceiling above serverless limits.
- You’re running a custom or fine-tuned model that isn’t in the catalog.
- You need a pinned model version for the long term.
Not sure which applies? and we’ll help you work it out.