Running a VLM on security camera feeds — what's the smallest model that won't ha

On March 22nd, an engineer posted a question to the LocalLLaMA community that most AI coverage would walk straight past: which vision-language model is small enough to run continuously on security camera feeds, and disciplined enough not to hallucinate when the footage is 720p night infrared — the kind of footage where a plastic bag catches a breeze and a model decides it's a crouching figure, where a cat's heat signature blooms into something that reads like a child. This wasn't a theoretical question about benchmark performance. It was a deployment question, the kind that only gets asked after someone has already burned through two or three candidate models and watched them confidently describe things that weren't there. The 720p night IR constraint is doing a lot of work in that sentence. It isn't an edge case — it's the standard operating condition for the majority of installed security cameras worldwide. The engineer wasn't asking about ideal conditions. They were asking about the real world, which is grainy and dark and full of visual ambiguity, and where a model's tendency to confabulate plausible-sounding descriptions isn't a quirk but a liability. The thread sits at a score of 7.0 in community engagement, suggesting the question landed with practitioners who recognised the problem immediately.

If confirmed as a genuine deployment challenge being actively worked through, this points to something the benchmark-driven AI conversation consistently underweights: hallucination in low-quality visual inputs isn't a research problem waiting to be solved at the frontier — it's a live operational problem for anyone trying to run vision models on commodity hardware against real-world footage right now. The security camera use case is particularly high-stakes because the downstream cost of a false positive isn't a mildly irritating chatbot response. It's a security guard dispatched to an empty car park, or worse, an automated system that escalates based on a model's confident misreading of shadow and thermal noise. The "smallest model that won't hallucinate" framing also reveals a real constraint that larger-model advocates tend to dismiss: continuous inference on camera feeds requires something that can run on edge hardware without melting the electricity budget, which immediately rules out the models that score best on standard vision benchmarks. There's a genuine gap between what the leaderboards measure and what this use case demands, and that gap is where practitioners are currently working without much guidance.

Watch for community responses naming specific models — particularly any smaller quantised variants of LLaVA, moondream, or similar lightweight VLMs — and whether anyone surfaces empirical results on infrared-specific hallucination rates rather than general visual benchmarks. That would signal the conversation is producing usable knowledge rather than just shared frustration.

NewsHive monitors these sources continuously. All signal titles above link to the original reporting.

Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →