An open video model aimed at lower long-input costs
Keye-VL-2.0-30B-A3B is a 30B-class open base model in the Keye family. It is built for understanding long videos and finding when specific events happen inside a video. Across five video tests, it is presented as leading other competitors and matching or beating Gemini-3-Flash on .
Its design uses , targeted feature gathering, , and other efficiency work so it can handle hour-long videos with less computation. The stack is also meant to cut the prefill cost for long inputs and improve training speed. The is curated to improve image understanding, OCR, charts, tables, and steady reasoning across steps.
Extra methods are used to make reasoning across text and visual information more reliable.
Key points
- It is a 30B-class open model focused on long-video understanding.
- It claims stronger video results than competitors across five benchmarks.
- and are used to reduce computation for long inputs.
- The model is designed to lower prefill cost and improve training throughput.
- Training focuses on OCR, chart and table understanding, visual reasoning, and continuity across steps.