A new company, Deep Cogito, has stealthily come out with a family of disposable models that can be changed between “reasoning” ways and not reasonable ways.
Reasoning models like O1 O1 have shown great promises in areas such as mathematics and physical, thanks to their ability to effectively control themselves by working through complex step -by -step problems. This reasoning comes at a cost, however: higher and latent computing. That is why laboratories as anthropic are following the architectures of the “hybrid” model that combine reasoning components with standard, not reasonable elements. Hybrid models can quickly answer simple questions as they spend extra time considering the most challenging questions.
All Deep Cogito models, called cogito 1, are hybrid models. Cogito claims that they exceed the best open models of the same size, including Meta models and the beginning of Chinese Deepseek.
“Each model can respond directly (…) or self-reflected before answering (as reasoning patterns),” the company explained in a blog post. “(All) were developed by a small team in approximately 75 days.”
Cogito 1 models range from 3 billion parameters to 70 billion parameters, and Cogito says models ranging up to 671 billion parameters will join them in the coming weeks and months. The parameters correspond to approximately the ability to solve the problem of a model, with more parameters generally better.
Cogito 1 was not developed from scratch, to be clear. Deep Cogito built on top of Meta’s Open Llama and Alibaba Qwen models to create his own. The company says it applied new training access to increase the performance of basic models and enable variable reasoning.
According to Cogito’s internal comparison results, the largest Cogito 1, Cogito 70b model, with the reasoning exceeds the R1 reasoning model of Deepseek in some assessments of mathematics and language. Cogito 70b with disabilities reasoning also eclipses the model of scout recently released Llama 4 in Livebench, a test with general goals.
Modelo Model Cogito 1 is available for downloading or use through APIs in Cloud fireworks providers he and together.
“Currently, we are still in the early stages of our scaling curve, using only a piece of reserved calculation typically for the traditional model of large language model/continuous training,” Cogito wrote in his blog post. “Speaking forward, we are investigating additional approaches after self-improvement training.”
According to registrations with the state of California, Deep Cogito based in San Francisco was founded in June 2024. Linkedin Page of the company order two co -founder, Drishan ARORA and Dhruv Malhotra. Malhotra was previously a Google Product Manager Ai Lab Deepmind, where he worked on search technology. ARORA was a high software engineer on Google.
Deep Cogito, whose supporters include South Park Commons, according to Pitchbook, aims ambitiously to build “general superintelligence”. The founders of the company understand the phrase which means one who can perform tasks better than most people and “discover completely new skills we have to imagine.”
