Plain-language definitions for the terms used across this repository.
Abliteration — A weight-level technique that reduces a model’s refusal direction without full retraining. Results vary by model and method.
API — Application Programming Interface. A contract that lets one program call another. In this repo, the contract is an HTTP endpoint that returns model completions.
API gateway — A service that sits between your app and one or more model providers. It handles auth, routing, billing and model selection.
Anthropic-compatible — An API that accepts the Anthropic Messages request shape (/messages, top-level system, max_tokens). Used by Claude Code and the Anthropic SDK.
Base URL — The root URL your SDK sends requests to. For OpenAI-style clients it often includes /v1; use the exact value from the RouterHub dashboard.
Cached input — Input tokens that a provider serves from a prompt cache. Usually billed at a lower rate than fresh input.
Chat completion — The OpenAI-style response format with a choices array. Used for conversational and instruction-following requests.
Completion — The model’s generated output for a prompt.
Context window — The maximum number of tokens a model can consider in one request, including input and output.
Function calling / tool use — A feature that lets a model return a structured call to an external function instead of, or in addition to, text. Support is model-dependent.
Inference — Running a trained model to produce output.
LLM — Large Language Model. A neural network trained on text to predict and generate language.
Model ID — The exact identifier you send in the model field. IDs change when providers release or retire versions.
Model family — A group of related models from one provider, such as Claude, GPT, Gemini, Grok, DeepSeek, GLM or Qwen.
OpenAI-compatible — An API that accepts the OpenAI chat completions shape (/chat/completions, messages array). Works with the OpenAI SDK and many third-party tools.
Pay per token — Billing based on the number of input and output tokens used, with no subscription.
Prompt — The text and instructions sent to the model.
Prompt caching — Reusing a previously processed prefix so repeated requests cost less and run faster.
Prompt injection — An attack where untrusted text tries to override the system’s instructions. Evaluate models for this before deploying agents.
Rate limit — A cap on requests or tokens over a time window.
Refusal — A model’s decision to decline or deflect a request. Uncensored variants reduce this behavior but do not eliminate it.
Streaming — Receiving a response incrementally as tokens are generated.
API key — A credential used to authenticate requests. Create separate keys for teammates, projects or clients.
System prompt — High-level instructions that set the model’s role and rules for a conversation.
Token — The unit models use to process text. Roughly, one token is a short word or part of a word; token counts vary by language and tokenizer.
Top-up — Adding balance to the account. RouterHub supports crypto top-ups.
Uncensored model — A model variant trained or tuned to refuse less often than a standard release. The term describes behavior, not a guarantee.