The Minimally Viable Language Model

As a software engineer, I’ve been diving back into AI, particularly small language models, after avoiding the hype (I’m a geek who hates overhype, not a hipster). I’m obsessed with the idea of a Minimally Viable Language Model (MVLM). Name aside, it’s about figuring out the smallest model that can still be useful.

LLMs are massive—hundreds of gigabytes—with tons of duplication. Since English is mostly defined (though it evolves), how many books or articles does a model need to “Google” the rest? What’s the size—2GB, 6GB, 10GB, 50GB? And how cheap could training an MVLM be?

Many LLMs now lean on “Deep Research,” essentially AI Googling from known sites. How far can this go? With Google’s search quality tanking, an MVLM could be a game-changer. Imagine every phone—Android first, maybe iPhone—running an MVLM locally, giving Siri-like assistants a privacy-first boost.

I’ve said on HN for years: AI won’t take off until it runs on your phone. Privacy concerns are real—some companies ban AI outright—and I’ve been downvoted for it, but the goalposts haven’t moved.