MirilAI/Miril-Drone-2B-1

VISIONConcurrent Unit Cost:1Model Size:5.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 7, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Miril-Drone-2B-1 is a 2 billion parameter aerial Vision-Language Model (VLM) developed by Miril.ai, based on Gemma4-2B. It is specifically designed for civilian drone-view imagery, enabling drones to reason about their environment and provide structured JSON responses. The model excels at scene description, visual question answering, and identifying operational points for review in aerial contexts, making it suitable for applications like delivery, inspection, and first response.

Loading preview...

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

Miril-Drone-2B-1 is a 2 billion parameter open-weight aerial Vision-Language Model (VLM) developed by Miril.ai, built upon the Gemma4-2B architecture. It is specifically engineered to process drone-view imagery, allowing drones to interpret and communicate about their surroundings. The model accepts an overhead image and a text prompt, returning structured JSON outputs for various tasks.

Key Capabilities

  • Drone-view scene descriptions: Provides compact, whole-frame descriptions of aerial images.
  • Aerial visual question answering: Answers questions about visible objects, counts, scene types, and potential hazards.
  • Operational pointing: Offers rough, image-relative coordinate cues for tasks like identifying safe landing zones, delivery areas, or inspection points. These are intended as review cues, not for autonomous flight control.
  • Structured JSON outputs: Generates machine-readable JSON responses for integration with operator tools and autonomy workflows.
  • WALDO Lineage: Incorporates aerial vocabulary from the WALDO perception line, enabling natural language responses for drone-relevant objects like LightVehicle, Building, and Person.

Use Cases

Miril-Drone-2B-1 is ideal for applications requiring small, edge-oriented VLMs to interpret drone footage. This includes aerial image review, drone operator assistance, field inspection triage, first-response situational awareness, and pre-screening of delivery or landing areas. It is a foundational model demonstrating future capabilities for making aerial systems easier to supervise and scale, though it is not yet a production-ready safety system.

Finetunes
6 models