Deep Cogito emerges from stealth with hybrid AI 'reasoning' models
Deep Cogito unveils powerful hybrid AI models for reasoning and real-time problem-solving.
Deep Cogito recently introduced a lineup of advanced AI models under the name Cogito 1, distinguishing themselves through the novel capability of transitioning between reasoning and non-reasoning modes. This approach was inspired by the pioneering work of models like OpenAI's o1, renowned for their structured reasoning skills, albeit with increased computational demands. To balance efficiency and processing power, Deep Cogito’s hybrid models can seamlessly shift between rapid responses for basic inquiries and more involved contemplation for complex scenarios. As a result, they stand to redefine performance expectations in the AI landscape, particularly in fields like mathematics and physics.
The Cogito 1 series, featuring models that start at 3 billion parameters and go up to 70 billion, promises further expansion in the upcoming months, aiming to reach 671 billion parameters. The parameter count is illustrative of a model's ability to address complex issues, and the company affirms that their models surpass the top-tier open models, such as those from Meta and DeepSeek. The foundational framework of Cogito 1 leverages Meta’s Llama and Alibaba’s Qwen models, with enhancements achieved through proprietary training techniques, enabling the activation and deactivation of reasoning faculties based on operational demands.
Drishan Arora and Dhruv Malhotra, the founders of Deep Cogito and former Google affiliates, spearheaded the development journey of Cogito 1. Both co-founders bring a wealth of experience from high-stakes tech environments: Malhotra's expertise in generative search at DeepMind, and Arora's engineering acumen, both crucial to the realization of their ambitious vision. Since its inception in June 2024, Deep Cogito, operating out of San Francisco with financial backing from South Park Commons, has pursued the bold goal of creating a 'general superintelligence' — an AI system performing tasks with a proficiency exceeding that of most humans.
Deep Cogito's models are readily available for integration and utilization, accessible via APIs through cloud services provided by Fireworks AI and Together AI. This explicit move towards accessibility aligns with their commitment to expanding the AI ecosystem, providing researchers and developers with cutting-edge tools to innovate continually. Cogito 1's largest model, Cogito 70B, has already demonstrated superior performance over the existing standards, including outshining DeepSeek's R1 in mathematics and language evaluations, as well as outperforming Meta's Llama 4 Scout on general AI assessment metrics such as LiveBench.
Looking ahead, Deep Cogito is focused on refining their scaling operations, having only scratched the surface of computational resources typically allocated for extensive model training. Their strategic pursuit of supplementary post-training methodologies aims to foster self-enhancing systems, driven by a commitment to exploring new frontiers of AI capabilities—effectively pushing the barriers of what machine intelligence can achieve.
Sources: Axios, TechCrunch, Deep Cogito Blog