Deep Cogito Launches Hybrid AI Models with Switchable Reasoning Modes
Deep Cogito, a new open-source AI company, has unveiled its first family of AI models that allow users to toggle between standard and advanced reasoning modes, as reported by TechCrunch. These models, named Cogito 1, are designed as hybrid systems that can swiftly handle straightforward tasks while engaging in deeper reasoning for more complex problems, such as intricate math or physics challenges.
This dual-mode approach aims to balance performance and efficiency, addressing the typical trade-off between computational cost and advanced reasoning abilities seen in purely reasoning-focused models.
Cogito 1 Models Promise Superior Performance
All Cogito 1 models, according to the company, surpass the performance of other open-source models of similar scale, including those developed by Meta and Chinese AI contender DeepSeek.
The Cogito 1 lineup spans a range from 3 billion to 70 billion parameters, with plans to release even larger models reaching up to 671 billion parameters shortly. In AI terms, a higher parameter count generally enhances a model’s ability to tackle complex tasks.
Built on Open Models, Enhanced with Proprietary Techniques
According to TechCrunch, rather than building entirely from scratch, Deep Cogito enhanced existing models from Meta’s Llama series and Alibaba’s Qwen models. Through novel training methods, the company claims to have significantly improved these base models’ capabilities and introduced the unique toggleable reasoning function.
Internal benchmarks suggest that Cogito 70B, the largest model currently available, outperforms DeepSeek’s R1 reasoning model in several math and language assessments. When its reasoning mode is turned off, Cogito 70B still outperforms Meta’s Llama 4 Scout in the general-purpose LiveBench evaluation.
Also read: Google Expands AI Mode Access with Enhanced Visual Search Features
Open Access and Cloud Availability
Demonstrating its commitment to openness, Deep Cogito has made all Cogito 1 models publicly accessible. Developers and enterprises can download the models directly or integrate them via APIs through cloud services such as Fireworks AI and Together AI.
