Introducing SSMoL: A Small Solar Model of Language
Our solar-powered AI assistant costs about $200 upfront and runs on literal sunshine. That’s roughly the price of ten months of a ChatGPT Plus plan, and after that it’s essentially free AI conversations forever. Like having a trust fund, but for talking to robots.
It’s “Free” Solar-Powered Small Language Model AI
(Yes, we could have called it a Small Solar Language Model, but then it wouldn’t spell SSMoL.)
How it works
Raspberry Pi 4 (8GB) running TinyLlama-1.1B-Chat — a lightweight AI model with 1.1B parameters
Solar battery pack for continuous 24/7 power
Wi-Fi connection with a simple API for sending prompts from a laptop
Prompting options: minimal web interface or straight from the command line
Total setup cost: about $200
Notes from Building
A small, locally run LLM is surprisingly useful. TinyLlama with 1.1B parameters handles 80% of what people actually use AI for. Turns out you don't need a nuclear reactor to fix typos or brainstorm pizza toppings. The research paper “Small Language Models are the Future of Agentic AI” backs up the idea that most AI applications don't need GPT-4’s full power. Small models are often the perfect first step for most AI tasks.
Your AI assistant could live where your vegetables do. Our solar-powered Pi sits in the backyard, literally. It's like having a shed with a very smart occupant who never sleeps, never needs feeding, and runs entirely on stored sunshine. The battery means it keeps chatting through the night, powered by yesterday's photons. Your conversations never leave your property. Not because of encryption or privacy policies, but because your AI physically lives in a box behind your house.
We achieved token neutrality. Every prompt costs tokens. Our tokens cost sunshine. ChatGPT shows a spinner and charges your credit card. Our Pi shows everything: “prompt eval time = 3074.16 ms / 37 tokens.” You can literally watch the AI think for free.
That haiku about renewable energy was written using renewable energy.
That email you needed cleaned up was edited with photons that traveled 93 million miles to fix your typos.
It’s like having a meter that always reads $0.00, no matter how many questions you ask.
Constraints make the product better… or maybe just give us perspective. The specs on our little Pi may sound tiny today, but it forces you to focus on what matters: a simple command-line interface to start, short fast prompts, and the kinds of tasks small models are actually great at.
And honestly, it took me back. I used to build giant desktop towers in my garage to run Cool Edit Pro with way less horsepower than this. I bought zip drives to back up PSD files and thought that was cutting edge. Now it’s 32GB on a chip the size of a thumbnail. Maybe that’s not a constraint after all. To butcher Bill Gates’s line: “32GB ought to be enough for anybody.”
This MVP ran for about 36 hours straight. On a sunny day the charging worked better than expected. A larger panel would take it further, but I was genuinely surprised it made it this far on a single charge.
Obviously this does not solve the problem. But as models get larger, more token hungry, and more “agentic,” it is worth asking: do we really need all of that? The future of AI might not be bigger models in bigger data centers. It might be smaller, smarter models… maybe even running on sunshine in your backyard.
