What we’ve shipped so far — research writing and open model releases. More landing soon.
Marlin-2B audits the standard dense-caption benchmarks (CaReBench, DREAM-1K) and finds 70%+ of the ground-truth captions are wrong about what is in the video. Then lands at #5 in open-sourceon TimeLens-Bench, beating GPT-5 at 3.5× smaller.
A small video VLM with two modes — Caption (structured scene and event description) and Find (timestamp-grounded retrieval). Competitive with Gemini-2.5 at 2B params. Use it out of the box or fine-tune it for your video agent use case.
Marlin 2B: tiny video model that slaps hard