MobileLLM-R1, a 950M-parameter model, matches Qwen3-0.6B on MATH500 and AIME while using 11.7% of its pretraining data. That didn't come from RL alone. It sits on a three-year stack: architecture, quantization, and data curation. This post traces that stack.

A 950M parameter model solves more competition math problems than models nearly twice its size. The gap isn't parameter count, it's training methodology and inference strategy. But cheap reasoning shifts the bottleneck to reliability: small models can reason, they just don't know when they're wrong.

Three years ago, running a language model on a phone meant a toy demo. Today, billion-parameter models run in real time on flagship devices. This shift came not from faster chips alone, but from rethinking how we build, compress, and deploy models.

A few years ago, watching teams throw hardware at AI’s memory problems, I started thinking: what if the bottleneck isn’t hardware, but how we use it?

There’s a growing consensus in AI that the next trillion-dollar platform won’t be another chatbot or copilot. It’ll be the system that captures context graphs: the decision traces, exceptions, and precedents that currently live in Slack threads, deal desk conversations, and people’s heads.