11/11/2026
Everyone talks about frontier models with hundreds of billions of parameters.
Pete from the EDGE AI FOUNDATION argues the real action is happening at the other end: small language models.
They keep getting better while getting smaller. A model with around a billion parameters fits in roughly a gigabyte of RAM, sometimes half that. That’s edge device territory.
And the pace is fast. Six months ago, small models like Qwen had context windows of about 8,000 to 12,000 tokens. Today it’s around 120,000. That’s roughly 10x in half a year.
In Pete’s words, models will be able to do a lot more in a lot smaller and more affordable spaces, every 6 months.
For anyone building connected devices, that changes the question. It’s no longer “can AI run on my hardware?” It’s “what should run locally, and what still needs the cloud?”
So where do you land? Is small, on-device AI ready for real products, or are frontier models still the only ones worth building on?
🎥 Full clip https://youtu.be/hdxTx-cYB9o