Model: the computation and the learned state

An AI model is a computation that maps inputs to outputs using a particular structure and state. In a neural network, the architecture defines the operations and the learned parameters determine how those operations behave.

The same architecture can hold very different skills after different training. Conversely, a familiar model name can refer to several checkpoints, precisions or fine-tuned variants.

Follow one request

For a language model, a tokenizer converts text to IDs, embedding lookup turns IDs into vectors, network layers transform those vectors, and an output projection produces scores for possible tokens. Decoding selects a continuation from those scores.

The model is central, but the complete application can also include retrieval, tools and a context-management loop. Those surrounding components change what the system can accomplish without necessarily changing the model's weights.

To reproduce a result, identify the checkpoint and matching configuration, preprocessing and generation settings. An architecture diagram alone does not specify the runnable system.