How TokenChange works

TokenChange connects people who run inference servers with people who want to call models. One person can be both: earn credits by serving requests, then spend them on models someone else hosts.

​

Models and endpoints

A model is an open-weights model in the catalog, such as a specific release of an open model family. Each model has a slug that you pass as model in API requests. An endpoint is one provider’s deployment of a model. A popular model can have many endpoints, each with its own prices, context window, quantization, capabilities and capacity. TokenChange doesn’t host model weights itself; every endpoint is run by an independent provider. When you call a model, TokenChange picks one of its endpoints for you. You can also pin a request to a single endpoint.

​

Providers

A provider is a user who hosts endpoints. Every new endpoint is reviewed by an operator and then validated automatically before it serves traffic. See Become a provider.

​

Credits

Credits are the only currency on TokenChange. Using an endpoint spends credits at that endpoint’s per-million-token prices. Hosting earns credits: providers receive what customers are charged on their endpoints, minus a 15% platform fee. See Credits and billing.

​

Inference keys and API tokens

TokenChange has two kinds of credential, and neither works in place of the other.

​

What happens to a request

1. You send a streaming request to inference.tokenchange.io with an inference key.
2. TokenChange checks the key and that your balance is above zero.
3. It finds the model’s endpoints that are serving, that your key allows and that fit any price options you set.
4. It routes the request to one of them. Requests that share a prompt prefix tend to reach the same endpoint, so the provider’s prompt cache is reused.
5. If that endpoint fails before sending anything, the request moves to another endpoint. You only see the stream from the endpoint that succeeds.
6. When the stream ends, the request is priced from its token usage and the endpoint’s prices, and the credits move from your balance to the provider.

Requests that deliver no output are never billed.

​

Next steps