Dongxi NLP explains how a language model produces each token

Dongxi NLP, a Chinese-language X account, published an explainer on how a chat model generates one token at a time.

Using the open-source Qwen3-0.6B model, the post traces the question "Is OpenAI open?" through a chat template, tokenization, prefill, sampling and detokenization. It says the answer takes nine tokens, and that each token requires a full model run that scores all 151,936 vocabulary entries.

The article links these steps to observed behavior, saying they explain why output appears gradually, why answers can vary, and why providers charge by token. A second post by the same account describes the same concepts, including logits, the KV cache, and the prefill and decode stages, as a basic overview rather than a new model or product.