Skip to content

GGUF version of MobileVLM V2 #70

Description

@sudarshan-krishnan

I'm currently building a mobile dashcam assistant that uses a vision-language model to evaluate whether the road ahead is safe for driving — detecting obstructions like vehicles, pedestrians, or debris in real time, and generating structured JSON outputs at ~2–3 frames per second.

I’ve been testing with SmolVLM 2.2 Q4_K_M inside PocketPal, which runs well, but I'm reaching limits in spatial reasoning and accuracy. I’d like to switch to MobileVLM V2 (1.7B or 3B) because:

It offers better multimodal grounding and occlusion handling
It’s explicitly optimized for mobile inference and real-time performance
It has strong benchmarks on scene understanding tasks (GQA, POPE, VQAv2, etc.)
However, MobileVLM V2 is currently not available in GGUF format and may require a modified projection layer (LDP V2) and prompt/token handling not currently supported by PocketPal.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions