← All models
Marlin is a 2B video VLM tuned for the two questions developers actually want to ask of their videos: what is happening, and when?
Details
- Accepts
- video
Pricing
- Input
- 1875 cr / 1M tokens
Prices in credits (1 credit = $0.01).
Data schema
Input
| Field | Type | Description |
|---|---|---|
| top_p | number | Nucleus sampling threshold. Only used when do_sample is true. |
| promptrequired | string | Natural-language description of the event or scene to find in the video. |
| do_sample | boolean | If true, sample with temperature/top_p; if false, use greedy decoding. |
| video_urlrequired | string | URL of the video to search through. |
| temperature | number | Sampling temperature. Only used when do_sample is true. |
Output
| Field | Type | Description |
|---|---|---|
| span | — | Detected (start, end) span in seconds, or null if the model could not locate the event. |
| text | string | Raw post-thinking model output (e.g. "From 14.3 to 18.2."). |
| format_ok | boolean | True iff the output matched the trained "From X to Y." format. |