Deep Cogito, a new AI company, has emerged from stealth mode, introducing a family of openly available AI models called Cogito 1. These models are designed with a unique hybrid architecture, allowing them to switch between “reasoning” and non-reasoning modes, potentially revolutionizing how AI handles complex tasks. Reasoning models, exemplified by OpenAI’s o1, have demonstrated remarkable capabilities in fields like mathematics and physics. Their strength lies in their ability to self-validate by meticulously working through problems step by step. However, this reasoning prowess comes at a cost: increased computational demands and latency. To address this, labs like Anthropic are exploring hybrid model architectures that combine reasoning components with standard, non-reasoning elements. These hybrid models offer the best of both worlds, quickly answering straightforward questions while dedicating more resources to intricate queries. Deep Cogito's Cogito 1 models are entirely hybrid, and the company asserts that they outperform the best open models of comparable size, including those from Meta and the Chinese AI startup DeepSeek. According to a company blog post, each model can either answer directly or engage in self-reflection before responding, similar to reasoning models. Impressively, these models were developed by a small team in approximately 75 days. The Cogito 1 models span a range from 3 billion to 70 billion parameters, with plans to introduce models reaching up to 671 billion parameters in the coming months. Parameters generally reflect a model’s problem-solving aptitude, with higher parameter counts typically indicating greater capabilities. It's important to note that Cogito 1 wasn't built from scratch. Deep Cogito leveraged Meta’s open Llama and Alibaba’s Qwen models as a foundation, applying innovative training methodologies to enhance performance and enable toggleable reasoning. Internal benchmarking results indicate that the largest Cogito 1 model, Cogito 70B, outperforms DeepSeek’s R1 reasoning model on several mathematics and language evaluations when reasoning is enabled. Furthermore, Cogito 70B, with reasoning disabled, surpasses Meta’s recently launched Llama 4 Scout model on LiveBench, a general-purpose AI test. All Cogito 1 models are available for download or use via APIs on cloud platforms like Fireworks AI and Together AI, making them readily accessible to developers and researchers. Deep Cogito's founders, Drishan Arora and Dhruv Malhotra, bring significant expertise to the company. Malhotra previously worked as a product manager at Google AI lab DeepMind, focusing on generative search technology, while Arora was a senior software engineer at Google. With backing from firms like South Park Commons, Deep Cogito aims to achieve “general superintelligence,” defined as AI that surpasses human capabilities and unlocks entirely new possibilities. The company acknowledges that they are still in the early stages of their scaling curve, having utilized only a fraction of the computational resources typically allocated for traditional large language model training. Looking ahead, Deep Cogito is exploring complementary post-training approaches for self-improvement, signaling a commitment to continuous innovation and pushing the boundaries of AI capabilities.
