A new AI startup, Deep Cogito, has officially launched out of stealth mode, introducing a new class of hybrid AI models designed to toggle between “reasoning” and non-reasoning modes. The company’s first model family, Cogito 1, offers openly available models that aim to outperform leading open-source alternatives, including Meta’s Llama and DeepSeek’s R1.
What Makes Cogito 1 Stand Out?
Reasoning AI models—like OpenAI’s o1—have gained attention for their ability to tackle complex tasks such as advanced math and physics by thinking step-by-step. But this capability usually comes with trade-offs in terms of higher computational costs and increased response times. That’s why researchers are now exploring hybrid model architectures, blending fast non-reasoning components with deeper reasoning logic that activates only when necessary.
Cogito 1 models follow this hybrid approach, offering users the flexibility to either receive immediate answers or allow the model to reflect before responding. According to Deep Cogito, this method allows for both speed and accuracy, depending on the complexity of the prompt.
Model Sizes and Performance Claims
The Cogito 1 series currently includes models ranging from 3 billion to 70 billion parameters, with much larger versions—up to 671 billion parameters—expected to roll out soon. Generally, a higher number of parameters translates to stronger problem-solving skills.
Despite building on existing open-source models like Meta’s Llama and Alibaba’s Qwen, Deep Cogito has reportedly introduced proprietary training techniques that enhance performance and enable the dynamic reasoning feature.
In internal testing, the largest current model, Cogito 70B, with reasoning enabled, outperforms DeepSeek’s R1 on several math and language benchmarks. Even with reasoning turned off, Cogito 70B beats Meta’s new Llama 4 Scout on LiveBench, a test designed to evaluate general-purpose AI capabilities.
Open Access and Cloud Availability
All Cogito 1 models are free to download or can be accessed via API through Fireworks AI and Together AI, two cloud platforms that support scalable AI deployment. The company has indicated that its current training process used only a small fraction of the computing resources typically consumed in large-scale AI training, leaving room for further optimization and post-training self-improvement strategies.
Who’s Behind Deep Cogito?
Based in San Francisco, Deep Cogito was founded in June 2024, according to public records. The company is led by co-founders Drishan Arora and Dhruv Malhotra. Malhotra brings experience from his time at DeepMind, where he focused on generative AI for search, while Arora previously worked as a senior software engineer at Google.
Their vision for Deep Cogito is bold: to build “general superintelligence”, which they define as AI capable of outperforming most humans across a wide range of tasks—and eventually revealing abilities that don’t yet exist.
Backed by early-stage investor South Park Commons, Deep Cogito positions itself as a serious contender in the race to develop AI models that are not only fast and accessible but also capable of deep reasoning when it matters most.
