Deep Cogito Presents Hybrid AI Models Capable of Reasoning and Direct Answering

San Francisco-based startup Deep Cogito has launched its first family of artificial intelligence models, called Cogito 1, which feature a hybrid architecture allowing them to toggle between ‘reasoning’ and non-reasoning modes. These models aim to provide efficient performance by addressing simple queries quickly, while taking more time to handle complex ones through step-by-step reasoning. Reasoning models have gained popularity for tasks in mathematics and physics due to their self-verification capabilities, though they typically require higher computational resources.

Cogito’s models range from 3 billion to 70 billion parameters, with plans for a 671 billion-parameter version soon. Built on top of Meta’s Llama and Alibaba’s Qwen models, Deep Cogito claims to have introduced novel training methods that enhance performance and allow mode switching. Internal benchmarks suggest Cogito’s 70B model surpasses DeepSeek’s reasoning model on specific tasks and even outperforms Meta’s latest Llama 4 Scout in general-purpose AI tests when reasoning is disabled.

Founded in June 2024 by former Google employees Dhruv Malhotra and Drishan Arora, the company is backed by South Park Commons. Their long-term goal is to build artificial general intelligence, or AGI, that can outperform humans across a range of tasks. The Cogito 1 models are openly available through platforms like Fireworks AI and Together AI.