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.
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.