mlx-bun library API
    Preparing search index...
    interface Qwen3VLTokenIds {
        imageTokenId: number;
        videoTokenId: number;
        visionEndId?: number;
        visionStartId?: number;
    }
    Index
    imageTokenId: number
    videoTokenId: number
    visionEndId?: number
    visionStartId?: number

    <|vision_start|> / <|vision_end|> — the per-frame-group wrappers the video expansion emits (defaults are the Qwen3.5 family ids).