Cogito v2

Cogito v2 is DeepCogito's second-generation family of open-weight hybrid reasoning models, released in mid 2025. It shipped in four sizes, 70B, 109B, 405B, and a 671B mixture-of-experts flagship. Each model can answer directly or switch into an internal reasoning mode when a question demands it.

The defining idea is Iterated Distillation and Amplification, a training loop that distills a model's own successful reasoning traces back into its weights. Over iterations the model develops what DeepCogito calls machine intuition: it gets better at guessing the right search trajectory, so reasoning chains come out roughly 60 percent shorter than rivals like DeepSeek R1 without losing accuracy.

DeepCogito reported training all eight Cogito v1 and v2 models for under $3.5 million total. A v2.1 revision followed later in 2025. The company positions the family as the strongest open-weight option from a US lab, competitive with frontier closed models on instruction following, coding, and long-context work. All weights are open source; benchmark figures cited are vendor-reported.