Deep Cogito Launches Hybrid AI Reasoning Models

Deep Cogito Launches Hybrid AI Models with Switchable Reasoning

A new AI company, Deep Cogito, has recently exited stealth mode, introducing a series of AI models that are publicly accessible. These models possess a unique capability: they can be dynamically switched between modes focused on “reasoning” and those operating without it.

The Advantage of Reasoning Models

Models exhibiting reasoning capabilities, such as OpenAI’s o1, have demonstrated significant potential in complex fields like mathematics and physics. This strength stems from their ability to rigorously verify their own conclusions through a step-by-step problem-solving process. However, this enhanced reasoning often necessitates greater computational resources and introduces increased latency.

Consequently, research groups like Anthropic are developing “hybrid” model architectures. These designs integrate reasoning components with conventional, non-reasoning elements. This approach allows for rapid responses to straightforward inquiries while dedicating more processing time to more intricate challenges.

Introducing Cogito 1

Cogito 1 represents Deep Cogito’s entire family of models, and all are designed as hybrid systems. The company asserts that these models surpass the performance of other openly available models of comparable size, including those originating from Meta and the Chinese AI firm, DeepSeek.

As explained in a company blog post, each model can provide a direct answer or engage in self-reflection before responding, mirroring the behavior of reasoning-focused models. Remarkably, the entire suite was created by a small team within approximately 75 days.

Model Sizes and Future Expansion

The Cogito 1 models currently range from 3 billion to 70 billion parameters. Generally, a higher parameter count correlates with improved problem-solving abilities. Deep Cogito anticipates the release of even larger models, up to 671 billion parameters, in the coming weeks and months.

Building on Existing Foundations

It’s important to note that Cogito 1 wasn’t built entirely from the ground up. Deep Cogito leveraged Meta’s open-source Llama and Alibaba’s Qwen models as a foundation. The company then applied innovative training methodologies to enhance the base models’ performance and enable the switchable reasoning feature.

Performance Benchmarks

Internal benchmarks conducted by Cogito indicate that the largest model, Cogito 70B, outperforms DeepSeek’s R1 reasoning model when reasoning is enabled, across several mathematical and linguistic evaluations. Furthermore, Cogito 70B, with reasoning disabled, achieves superior results compared to Meta’s recently launched Llama 4 Scout model on LiveBench, a comprehensive AI testing platform.

Accessibility and Availability

All Cogito 1 models are readily available for download or access through APIs hosted on the cloud platforms Fireworks AI and Together AI.

Future Development

In a blog post, Cogito stated that they are still in the initial phases of scaling their operations, having utilized only a small fraction of the computational power typically allocated to large language model training. They are actively exploring additional post-training techniques to further enhance self-improvement.

Company Background

According to filings with the state of California, Deep Cogito was founded in June 2024 and is based in San Francisco. The company’s LinkedIn profile identifies two co-founders: Drishan Arora and Dhruv Malhotra.

Malhotra previously held a position as a product manager at Google AI’s DeepMind, where his work centered on generative search technology. Arora was a senior software engineer at Google.

Ambitious Goals

Backed by investors including South Park Commons, Deep Cogito has set an ambitious goal: to develop “general superintelligence.” The founders define this as AI capable of exceeding human performance in most tasks and discovering entirely new capabilities beyond our current imagination.