A new company Deep Cogito () has emerged from the shadows with a family openly available AI model that can be switched to “reasoning mode” or non-reasoning mode.
OpenAI’s o1 and other reasoning models have shown great promise, especially in math and physics domains. This is because they can fact-check by themselves by working through complicated problems step by step. However, this reasoning comes with a price: higher computing and a longer latency. Anthropic, for example, is pursuing “hybrid model architectures” that combine reasoning elements with standard non-reasoning components. Hybrid models are able to quickly answer simple queries while spending more time on more complex queries.
Deep Cogito models, including Cogito 1, all are hybrid models. Cogito claims they outperform open models of similar size, such as models from Meta and Chinese AI start-up DeepSeek.
Each model can answer directly or self-reflect (like reasoning models), according to the company. In a blog postit is explained. “[All] was developed by a small group in approximately 75 days.” (19659004) The Cogito 1 model ranges from 3 billion to 70 billion parameters. Cogito states that models up to 671 trillion parameters will be joining them in the next weeks and months. The number of parameters roughly corresponds to the model’s ability to solve problems. More parameters are generally better. To be clear, Cogito 1 was not developed from scratch. Deep Cogito created its own models by building on top of Meta’s open Llama model and Alibaba’s Qwen model. The company claims that it used novel training methods to boost the performance of base models and enable toggleable reason.
According the results of Cogito’s internal benchmarking the largest Cogito 1 Model, Cogito70B, with reasoning, outperforms DeepSeek’s R1 reasoning on a few math and language evaluations. Cogito 70B without reasoning also outperforms Meta’s Llama 4 Scout, a recently released general-purpose AI model.
Each Cogito 1 model can be downloaded or used via APIs at cloud providers Fireworks AI, and Together AI.
“Currently, we’re still in [our] the early stages of scaling curve, after having used only a small fraction of compute typically reserved to traditional large language model posts/continued learning,” wrote Cogito on its blog. “Moving ahead, we’re exploring complementary post-training methods for self-improvement.” According to filings at California StateDeep Cogito, based in San Francisco, was founded on June 20, 2024. The company’s LinkedIn page shows two co-founders: Drishan Arora, and Dhruv malhotra. Malhotra worked as a product manager for Google AI lab DeepMind where he developed generative search technology. Arora was previously a senior software developer at Google.
Deep Cogito’s backers include South Park Commons. According to PitchBookthe company aims to build a “general superintelligence” that can perform tasks more efficiently than humans. It also aims to “discover entirely new capabilities which we have not yet imagined.”
Kyle Wiggers, TechCrunch AI Editor. His writings have appeared in VentureBeat, Digital Trends and a variety of gadget blogs, including Android Police and Android Authority. He lives in Manhattan, with his music therapist partner.
View Bio
