Video understanding that runs in real time, on the edge
Most video AI today relies on vision-language models, which interpret footage by predicting text tokens from labeled frames. That approach is expensive to run, usually tied to the cloud, and prone to hallucination. Primate Intelligence is building on a different foundation: predictive world models based on JEPA (Joint Embedding Predictive Architecture), which learn directly from video rather than from text labels.
Primate's first model, Darwin, recognizes actions and objects in live streams or uploaded video. Users ask questions in plain language and can change the prompt mid-stream, with no per-use-case training or labeling. Darwin is available through the Primate Vision app and a developer API, and the company is building an SDK for edge hardware including NVIDIA Jetson, Mac, iOS, and Snapdragon. Primate positions it as delivering cloud-level benchmark accuracy at roughly a tenth of the cost of VLM-based approaches.
Because Darwin's outputs are deterministic and grounded in what it observed, Primate is targeting settings where errors are costly and latency matters: smart cameras, robotics, drones, autonomous vehicles, healthcare, and video search.