MirilAI/Miril-Drone-2B-1

Miril-Drone-2B-1: Aerial Vision-Language Model

Miril-Drone-2B-1 is a 5.1 billion parameter open-weight aerial Vision-Language Model (VLM) developed by Miril.ai, specifically designed for drone-view imagery. Built on Gemma4 multimodal components, it enables drones to interpret and communicate about their environment by processing overhead frames and instructions to return structured JSON responses.

Key Capabilities

• Aerial Visual Understanding: Provides drone-view scene descriptions, visual question answering, and ground-context review.
• Structured Outputs: Generates bare JSON outputs for integration with operator tools and autonomy workflows, including captions, answers, and operational coordinates.
• Operational Pointing: Supports rough representative grid cues for tasks like identifying candidate landing areas or inspection points, though it's not a certified safety system.
• WALDO Lineage: Inherits aerial vocabulary from the WALDO perception line, enabling it to describe objects like LightVehicle, Building, Person, and UPole in drone imagery.

Intended Use Cases

• Drone Operator Assistance: Aids in aerial image review, field inspection triage, and first-response situational awareness.
• Pre-screening: Useful for delivery or landing-area pre-screening.
• Research: Supports research into aerial VLMs and edge-oriented perception.

Limitations and Safety

This V1 model is a demonstration and not a production-ready safety system. It can miss small objects, confuse scale, and its pointing outputs are rough cues, not precise localization. It is intended as a component in a larger stack with human oversight and independent safety checks. Variants are available for different deployment paths, including CUDA and Apple Silicon MLX.