Temps réel, flux, lot, différé, appareil : les cinq façons de servir…
A trained model is useless until someone can ask it something. How you make it available comes down to five modes, and what separates them is not technology but one plain question: who is waiting for the answer, and for how long? **Real time**: one request, an answer within milliseconds, and a human waiting at the screen, like the check on a payment the moment you click pay. **Streaming**: the same model endpoint, fed by an event queue instead of a click, like a delivery arrival time recomputed every time the courier moves. **Batch**: predictions computed for everyone at once on a schedule, like a churn risk score recalculated every night for every subscriber. **Asynchronous**: work too heavy to wait for, which you submit and get notified about when it is done, like summarising a long PDF; on AWS this mode takes payloads up to 1 GB and up to an hour of processing. **On device**: the model runs on the phone or the watch, with no network, no round trip to a server and no data leaving the user, which is why small language models get so much attention. Large language models split the same way: at Anthropic, the batch processing API costs half the price of a regular call, and most batches finish in under an hour.