LLM API glossary

Large language model: a neural network trained on large amounts of text that predicts the next token, which lets it answer, summarize, translate or write code.

A model whose trained parameters are published, so others can run it on their own hardware. What they may do with it depends on its licence.

A model only its maker serves, such as GPT, Claude, Gemini or Grok. Access goes through the maker or partners it approves.

The terms under which open weights are published: Apache 2.0 and MIT allow commercial serving; the Llama community licence adds conditions such as a use policy and the “Built with Llama” notice.

The numbers inside a model. “120b” means about 120 billion; more parameters usually means more capable and more expensive to run.

A design where only part of the model works on each token. In “Qwen3 235B A22B”, 235 billion parameters exist and about 22 billion are active per token.

A model trained to work through a problem step by step before answering, such as DeepSeek R1. It often spends more output tokens per answer.

Storing a model's numbers with fewer bits so it runs on less memory. It can lower quality slightly; hosts do not always say which precision they serve.

The maximum number of tokens a model can take in at once, prompt and answer together. Ours range from 128K to 1M tokens.

An API that accepts the same request and response format as OpenAI's, so its SDKs and the tools built on them work after changing the base URL.

The address an SDK sends requests to. For LLMwithCrypto it is https://llmwithcrypto.com/api/v1.

The OpenAI request format with a list of messages (system, user, assistant) and a reply from the model. Our main endpoint is https://llmwithcrypto.com/api/v1/chat/completions.

A secret string that identifies your account on each call, sent as a Bearer token. Keep it in an environment variable, never in code.

Receiving the answer token by token as it is generated, through server-sent events, instead of waiting for the whole reply.

Letting the model ask your code to run a function, with arguments in JSON, and then use the result. Also called function calling.

Asking the model to answer with valid JSON, set through response_format on models that support it.

The number of requests or tokens an account may send per minute. Ours is shown in the dashboard and can be raised on request.

The first message of a conversation that sets the model's role and rules for the rest of it.

The unit models read and write: a word or part of a word. In English, 1,000 tokens are roughly 750 words.

Input tokens are what you send, output tokens what the model writes. Output usually costs more per token.

How model APIs quote prices: the cost of one million input or output tokens. A call costs a fraction of it.

The rule a model uses to cut text into tokens. The same text counts a different number of tokens on different models.

The middle value of the prices that public hosts list for the same model. Our prices are 5% below it.

A discount some hosts give when the start of a prompt repeats. We do not offer it yet.

Prepaid API credit bought at face value from your balance; it pays for any model and never expires.

Know your customer: checks of identity documents or company papers before a service can be used. LLMwithCrypto does not do them.

The token standard of the Tron network. USDT on TRC-20 often costs the least to send for a small dollar amount.

The token standard of Ethereum, used for USDT and USDC there. Network fees can be high for small transfers.

What the blockchain charges to process your transfer, paid by the sender to the network, not to us.

A block added after your transaction. We credit a top-up once its network has the confirmations it needs.

The smallest amount you can add to your balance: $25.

The dollar amount on your account after a top-up, from which you buy credit packs.