DeepSeek has published a guide explaining how tokens are used as the basic unit for both text representation and billing in its models. Tokens can be thought of as characters or words, with each Chinese word, English word, number, or symbol typically counting as one token.
The company provides approximate conversion ratios: one English character equals about 0.3 token, while one Chinese character equals about 0.6 token. However, these ratios can vary depending on the tokenization method used by different models.
DeepSeek notes that the actual number of tokens processed per request is determined by the model's response, which users can check in the usage results. To help developers calculate token usage for their inputs and outputs, DeepSeek offers a demo tokenizer code in a downloadable zip package.