Changelog

​

Breaking

• The model no longer reports the prompt text. state drops current_prompt, and chunk_complete drops active_prompt. Keep the prompt your app sent if your UI shows it. prompt_accepted still echoes the text you sent.

​

New

• set_prompt takes a passthrough flag. The deployment prepares your text on the server before it conditions generation. Send passthrough: true to skip that step, and the model gets your text unchanged. Use it for a prompt you already wrote against the model’s own caption schema. The flag applies per call, and the default is false, so a client that never sends it keeps the behavior it had.

​

Changed

• A run reaches 2000 chunks (~61 minutes) rather than 229 (~7 minutes). generation_complete still returns the session to WAITING at the cap, and the model does not start again on its own. Expect quality to drift on a long run, well before the cap.