The Sequence AI of the Week #769: Inside Gemini Deep Think
One of the most innovative AI architectures of the last few years.
Gemini Deep Think is one of the most innovative architectures of recent times and, yet, we know so little about it. Today, I would like to summarize some of the things I learned about Deep Think.
Gemini DeepThink made news when it score a gold medal at the 2025 international math olympiad using a parallel technique over the standard Gemini model. Deep Think sits at an interesting point in the evolution of large-scale language models: it’s not a brand-new backbone, but a “thinking layer” built on top of Google’s Gemini 2.5 and now Gemini 3 architectures that turns a big multimodal MoE into a coordinated swarm of reasoning agents. It embodies the current frontier idea that how a model uses its compute at inference time matters as much as raw parameter count.
Below is a technical look at how we got here, what the DeepThink architecture actually is (as far as Google has disclosed), and why its results on Olympiad math, competitive programming, and frontier benchmarks matter for the rest of the field.