News Summary

Deep Cogito, a new AI company, has emerged from stealth, introducing a series of hybrid AI models known as Cogito 1, which can switch between reasoning and non-reasoning modes. These models are designed to outperform existing open models, including those from Meta and Chinese startup DeepSeek. Developed in just 75 days by a small team, Cogito 1 models range from 3 billion to 70 billion parameters, with plans for larger models on the horizon. Built on Meta’s Llama and Alibaba’s Qwen models, Cogito 1 utilizes innovative training methods to enhance performance. The largest model, Cogito 70B, with reasoning capabilities, surpasses DeepSeek’s R1 in mathematics and language tests and also outperforms Meta’s Llama 4 Scout model when reasoning is disabled. Available via APIs on platforms like Fireworks AI and Together AI, Deep Cogito aims to eventually build general superintelligence. Founded in June 2024 by former Google employees Drishan Arora and Dhruv Malhotra, the company is backed by South Park Commons.