Image, video, speech and music generation

Image, video, speech and music each have one global selection under Settings → Multimodal Models; the available sources, reference images, per-source limits, and how generated files appear in the session.

Image generation, speech, music, and video generation each have one global selection under Settings → Multimodal Models. A source's sign-in state decides which options are available; when a request actually runs, only the source selected here is used. All chats share these selections — a chat using Claude, Codex, or another language model does not change which source generates.

In the current version, generation runs only on subscription sources you have connected; there is no custom-endpoint option. Each source uses that account's own sign-in credentials, with no separate backend to fill in.

The generation tools are always callable, and every request reads the current selection. When the selected source is not signed in or the request fails, the tool returns an explicit error; it never quietly switches to another configured source. After you change the selection, subsequent requests use it — there is no need to switch the chat's language model.

Generation has a dedicated display in chat: image, video, audio and music each carry their own icon, and the status row follows the progress — from awaiting approval to generating, then generated, failed or stopped. Several generations in one message are grouped together, and any failure within the group is shown as it is.

Sources per capability

  • Image: Codex, Cursor, Qoder, QwenWork, WorkBuddy, Antigravity, TRAE CN, Bailian Token Plan, MiniMax M Plan.
  • Speech: Cursor, MiniMax M Plan, Bailian Token Plan, WorkBuddy.
  • Music: QwenWork, with optional lyrics.
  • Video: MiniMax M Plan, WorkBuddy, QwenWork, TRAE CN.

Per-source limits

  • Reference images (image editing): Codex, Cursor, Antigravity, TRAE, WorkBuddy, and QwenWork accept reference images; Qoder, Bailian, and MiniMax are text-to-image only. Each reference image is capped at 4 MiB, and 8 MiB in total per request. QwenWork image prompts are capped at 2500 characters.
  • MiniMax video requires the Explore or Build tier; Go is not supported. It keeps the video model name you set, but the endpoint and credentials always come from the MiniMax subscription.
  • Bailian speech goes through the plan gateway's own speech synthesis route and spends plan allowance; it accepts only qwen-audio-3.0 series models and returns a download link valid for 24 hours. Zhipu Coding Plan and Ark Coding Plan are not offered as speech sources.
  • MiniMax shares one subscription allowance across text, image, speech, and video.

Each speech request can also set a voice, emotion, and speed for that clip without changing the global settings: voice applies to every source; emotion (calm, happy, and so on) is honored only by MiniMax, and Cursor rejects emotion overrides; speed ranges from 0.75 to 2 times and is honored by the sources that support it. Unsupported parameters are ignored.

Local image, video, and audio generation is not available in the current version. A retained legacy local configuration does not block another explicitly selected subscription source; selecting that legacy configuration shows an unavailability notice.

Generated files count as artifacts of the turn, are saved inside the session directory, play inline in the conversation, and can also be opened from the Files panel — see Session artifacts.