Dedicated capacity
Dedicated Capacity overview
Inference capacity reserved for you in Munich: GLM-5.3, MiniMax M3.1 Flash and more on an OpenAI-compatible endpoint of your own. How it works and how to start.
What Dedicated Capacity is
Section titled “What Dedicated Capacity is”Dedicated Capacity is inference capacity reserved for one customer. It runs on purpose-built AI chips for inference, SambaNova’s Reconfigurable Dataflow Units (RDUs), in our data center in Munich. We run it together with our operating partner SambaNova, which operates the software stack with us. You call it through the same OpenAI-compatible API as the Inference API, on an endpoint of your own.
Models
Section titled “Models”- GLM-5.3 and MiniMax M3.1 Flash
- gpt-oss-120b, Gemma 4 and more. See Supported models.
- Your own fine-tuned checkpoints. See Custom checkpoints.
Model bundles
Section titled “Model bundles”You decide which models run on your capacity. Your capacity is one rack or several racks. A rack can run one model, or several models together in one bundle.
- Several models on one rack. The models in a bundle stay loaded together. The rack computes one model at a time and switches between them in milliseconds, per request. The models share the throughput of the rack, so we size a bundle for the combined traffic of its models. Each request selects its model by the model ID, on the same endpoint. For example, one rack can serve a text model, an embedding model and a speech-to-text model for a complete pipeline.
- Several configurations of one model. A bundle can be set up for your workload, for example for longer context lengths, for more concurrent requests, or with prompt caching.
- Your split across racks. With several racks, you decide how they are split between your models and bundles. We can change the split later.
Which models and configurations fit together depends on the models. We agree the specific bundles with you as part of the project.
Inference API or Dedicated Capacity
Section titled “Inference API or Dedicated Capacity”| Inference API | Dedicated Capacity | |
|---|---|---|
| Infrastructure | Shared infrastructure | Reserved for you |
| Model selection | Set by Infercom | Chosen by you |
| Models | The Inference API catalog | GLM-5.3, MiniMax M3.1 Flash and the full dedicated list, plus your checkpoints. Several models in one bundle, depending on the models |
| Access | api.infercom.ai, self-serve key from cloud.infercom.ai | An endpoint of your own, set up with you |
| Billing | Pay per token | Monthly, per agreement |
| Rate limits | By tier | Agreed with you |
| Throughput | Shared with other customers under load | Reserved for your traffic |
| Start | Minutes | Lead time on request |
When Dedicated Capacity fits
Section titled “When Dedicated Capacity fits”- Sustained production volume.
- Models that are not in the Inference API catalog.
- Your own fine-tuned checkpoints.
- Traffic that must not share capacity with other customers.
For variable or exploratory workloads, the Inference API is usually the better start. You can move to Dedicated Capacity later. The API stays the same.
How to start
Section titled “How to start”Talk to us. We size the capacity from your workload. See what to send us.
The lead time is available on request. After onboarding, see Connect to your endpoint.
The same models can also run on hardware in your own data center. See On-premises.