Deep Cogito, a new AI startup, has emerged from stealth mode with a groundbreaking family of hybrid AI models that can toggle between "reasoning" and non-reasoning modes. This innovative approach allows the models to balance computational efficiency with advanced problem-solving capabilities, making them a significant advancement in the field of artificial intelligence.

The Cogito 1 models, ranging from 3 billion to 70 billion parameters, have been developed by a small team in just 75 days. What's remarkable is that these models outperform some of the best open models of the same size, including those from industry giants like Meta and Chinese AI startup DeepSeek. This achievement is a testament to the company's novel training approaches, which have enabled the creation of hybrid models that can adapt to different problem-solving scenarios.

Hybrid models like Cogito 1 are particularly interesting because they combine the strengths of reasoning models, which can effectively fact-check themselves by working through complex problems step by step, with the speed and efficiency of non-reasoning models. This allows them to quickly answer simple questions while spending additional time considering more challenging queries. Anthropic, another prominent AI lab, is also pursuing similar hybrid model architectures, highlighting the growing importance of this approach in the AI research community.

Deep Cogito's models weren't developed from scratch, but rather built upon existing open-source models from Meta and Alibaba. The company applied its novel training approaches to boost the base models' performance and enable toggleable reasoning. This approach has paid off, with the largest Cogito 1 model, Cogito 70B, outperforming DeepSeek's R1 reasoning model on several mathematics and language evaluations. Additionally, Cogito 70B with reasoning disabled also eclipses Meta's recently released 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 providers Fireworks AI and Together AI, making them accessible to developers and researchers worldwide. This open approach is likely to accelerate the development of AI applications and foster further innovation in the field.

Deep Cogito's ambitious goal is to build "general superintelligence," which the company's founders define as AI that can perform tasks better than most humans and "uncover entirely new capabilities we have yet to imagine." With its innovative hybrid models, the company is taking a significant step towards achieving this vision. Backed by investors like South Park Commons, Deep Cogito is well-positioned to continue pushing the boundaries of AI research and development.

As the AI landscape continues to evolve, the emergence of Deep Cogito and its hybrid models is a significant development that warrants attention. With its focus on toggleable reasoning and hybrid architectures, the company is poised to make a meaningful impact on the industry. As AI continues to transform various aspects of our lives, innovations like Cogito 1 will play a crucial role in shaping the future of artificial intelligence.