Our benchmark suite covers three capability dimensions: action detection (does the model correctly identify that a given action occurred?), temporal localization (does it know when — with timestamps?), and question-answering with timestamps (can it answer open-ended questions about video content and cite the evidence?). We run these against YOLO-class detectors, VLM-based approaches (GPT-4o, Gemini), and the V-JEPA 2.1 family — the most honest baselines for each capability class.
Darwin beats V-JEPA 2.1 on SSv2 and EK-100 action benchmarks at a fraction of the training cost. On temporal localization tasks it outperforms every VLM we have tested, because VLMs sample frames — Darwin watches video. Honest failure modes are part of the published results: low-contrast scenes and heavily occluded objects remain harder for Darwin than for specialized detectors.
Extended methodology and next model release results — coming during our August launch.