# Querying a Fireworks deployment uses a different model string, not a different endpoint On-demand deployments on Fireworks do not get their own endpoint. You query them through the same OpenAI-compatible endpoint, https://api.fireworks.ai/inference/v1, but the model field changes: instead of accounts/fireworks/models/[name] you pass accounts/[account]/deployments/[deployment-id]. Same SDK, same /chat/completions path, different model string. Two money details from the same guide: - Aborting a stream means closing the connection. Do that when the user walks away: it stops generation, so you avoid billing for tokens that nobody will read. - To completely omit the instructions, set the first message's content to an empty string. Deleting the system message from your history array can change behavior in some renderers; the empty string is the documented way to say no instructions. The failure mode this prevents: hardcoding the serverless model id everywhere, then deploying your fine-tuned model and wondering why traffic never reaches it. Point the model field at the deployment identifier.

Context: Fireworks querying text models: deployment identifier format, stream abort billing Fireworks' text models guide documents querying dedicated deployments with the same OpenAI-compatible APIs, but with a deployment-specific identifier: model='accounts/[account]/deployments/[deployment-id]' at base URL https://api.fireworks.ai/inference/v1. Two billing details: aborting a stream means closing the connection, and closing it stops generation so you avoid billing for ungenerated tokens. Also note the guide: to omit the instructions entirely, set the first message's content to an empty string rather than dropping the system message.