The new company, DEEP COGITO, came from stealth with an open AI model family that can be transformed between “reasoning” and non -class mode.
The reasoning model, such as Openai’s O1, has shown great promises in areas such as mathematics and physics, thanks to the ability to effectively identify themselves by working with complex problems step by step. However, these reasoning is expensive: higher computing and waiting time. That’s why laboratories like Anthropic pursue a “hybrid” model architecture that combines reasoning components with standard non -class elements. The hybrid model quickly answers simple questions and considers more difficult queries to spend more time.
The model of Dip Kogi, called Cogito 1, is a hybrid model. COGITO claims to be better than the best open model of the same size, including the model of the meta and China AI startup DeepSeek.
The company has been developed by a small team in about 75 days in the blog post, “Each model can be answered directly before answering (…) or self -reflected.” (All) in about 75 days.
The range of the Cogito 1 model is from 3 billion parameters to 70 billion parameters, and COGITO says that up to 670 billion parameters will be combined for the next few weeks and months. Parameters are roughly the problem -solving techniques of the model, and more parameters are generally better.
Cogito 1 has not been developed from the beginning and is clear. Deep Cogito created its own by producing META’s Open llama and Alibaba QWEN models. The company has applied a new training approach to improve the performance of the basic model and enable the inference that can be used.
According to Cogito’s internal benchmarking results, the largest Cogito 1 model, Cogito 70B, is better than DEEPSEEK’s R1 reasoning model for some mathematics and language evaluation. COGITO 70B, which has an reasoning disorder, is also a general purpose AI test in Livebench, META’s recently released LLAMA 4 scout model.
All Cogito 1 models can be downloaded or used through the API of the cloud provider.
Cogito said in a blog post, “Currently, we are still in the early stages of the scaling curve, and we have used only a part of the general reserved computing for traditional large language model posts.
The San Francisco -based Deep Cogito, which was submitted to California, was founded in June 2024. The company’s LinkedIn page shows two co -founders, Drishan Arora and DHRUV Malhotra. Malhotra was a product manager of Google AI Lab Deepmind and worked in the creation search technology. Arora was a senior software engineer in Google.
According to the Peachbook, Deep Cogito, which includes South Park Commons, aims to build a “general super intelligence”. The founders of the company means AI, which can perform better than most humans, and understands the phrase, “We do not find completely new abilities we have not yet imagined.”
