Cogito v2.1 (Reasoning) is a large-scale language model developed by Deep Cogito, a San Francisco-based AI research lab. Released in November 2025, the model utilizes a 671B parameter Mixture of Experts (MoE) architecture and is designed to handle complex logical, mathematical, and coding tasks through a hybrid reasoning approach. It can switch between a standard direct-response mode and an internal self-reflection mode to verify intermediate steps before providing a final answer.
The model's core innovation is its training methodology, Iterated Distillation and Amplification (IDA). This strategy involves the model internalizing its reasoning processes through iterative self-improvement and process supervision. By distilling successful reasoning paths back into the model weights, Cogito v2.1 aims to achieve high-level cognitive performance with shorter internal chain-of-thought sequences compared to models that rely solely on lengthy inference-time search.
Technical Capabilities
Cogito v2.1 is optimized for technical domains including STEM, instruction following, and multilingual tool calling across more than 30 languages. It supports a 128k context window and is engineered to minimize token usage while maintaining accuracy in complex problem-solving. The model's "machine intuition" is trained to prune inefficient thought trajectories, resulting in faster and more efficient reasoning outputs than its predecessors.