Benchmarks & research
OneJev is an open-weight multimodal decision model family that evaluates text, images, and video to answer typed questions in a single forward pass.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
The training dataset and fine-tuning configurations are public. OneJev fine-tunes Qwen and handles all questions about a state at once, forking the cache after their shared context. Its supplied server accepts TypeSafe-compatible requests through PyTorch or llama.cpp. Only the PyTorch backend handles video.
Choose a model your hardware can run, then use the server to ask questions about a situation and any supported images. You specify the fixed answers; the server scores them without writing a reply. For video, use its PyTorch software rather than the lighter llama.cpp option, which cannot handle video questions.