Deep Cogito has emerged from stealth mode, unveiling a new family of openly accessible hybrid artificial intelligence models that can switch seamlessly between reasoning and standard modes. This approach addresses the major drawback of purely reasoning-oriented models, which often display high computational demands and latency.
Known as Cogito 1, the company’s models demonstrate the capability to efficiently handle straightforward inquiries while automatically activating more in-depth reasoning skills for complex problems. This dual approach helps maximize performance and efficiency, enabling practical adaptations across different tasks and computational environments.
Cogito models range in scale from approximately 3 billion to 70 billion parameters, with plans to expand to an even larger 671 billion parameter model in the coming months. The number of parameters typically influences an AI model’s problem-solving capabilities, with higher parameters often correlating to enhanced performance.
Built on foundations from Meta’s open-source Llama models and Alibaba’s Qwen models, Cogito integrates unique training methodologies. According to the company’s internal benchmarking results, its 70-billion-parameter flagship model, called Cogito 70B, outperformed competing models including DeepSeek’s R1 and Meta’s recent Llama 4 Scout model. Specifically, Cogito 70B achieved higher scores in mathematical reasoning tests and language comprehension evaluations, as well as on LiveBench—a widely regarded general-purpose AI assessment.
Deep Cogito stated that the models were developed by a relatively small team in about 75 days. Currently, Cogito models are accessible both as downloadable code and through APIs via cloud platforms such as Fireworks AI and Together AI.
Despite the rapid timeline, the startup emphasizes that this release is part of its initial steps in a longer journey toward continuous scaling and improvement. Its researchers are actively exploring new strategies for post-training enhancement intended to further escalate performance capabilities.
Founded in June 2024 and headquartered in San Francisco, Deep Cogito was co-founded by Dhruv Malhotra, a former product manager at Google’s DeepMind, and Drishan Arora, formerly a senior software engineer also at Google. The company, which counts South Park Commons among its early investors, states ambitious aims to eventually develop “general superintelligence,” or artificial intelligence systems capable of outperforming humans in a broad array of tasks and potentially demonstrating entirely new cognitive abilities beyond current imagination.
