The new company, Deep Cogito, has emerged from stealth along with a family of openly available AI models that can switch between “inference” and irrational modes.
Inference models like Openai’s O1 show great expectations for domains such as mathematics and physics, thanks to their ability to effectively fact-check themselves by working on complex problems in stages. However, this inference is costly. Higher computing and latency. So labs like humanity pursue “hybrid” model architectures that combine inference components with standard irrational elements. Hybrid models allow you to quickly answer simple questions while spending extra time considering more challenging queries.
All deep Cogito models, known as the Cogito 1, are hybrid models. Cogito claims it is better than the best open models of the same size, such as the models from Meta and Chinese AI startup DeepSeek.
“Each model can be directly (…) or self-reflected before answering (like an inference model),” the company explained in a blog post. “(Everything) was developed by a small team in about 75 days.”
The range of the Cogito 1 model ranges from 3 billion to 70 billion parameters, and Cogito says models with a range of up to 67.1 billion parameters will be participating in them in the coming weeks and months. The parameters are roughly compatible with the model’s problem-solving skills, and in general, more parameters are better.
Cogito 1 was not developed from scratch. Built on top of Meta’s open llama and Alibaba’s Qwen model, Deep Cogito created its own. The company says it has applied a new training approach to improve performance of the base model and enable toggable inference.
According to results from Cogito’s internal benchmarks, the largest Cogito 1 model, the Cogito 70b, surpasses Deepseek’s R1 inference model for several mathematics and language ratings. The Cogito 70b is a generic AI test with Eclipses Meta’s recently released Llama 4 Scout Model on LiveBench with Reasoning Disabled.
All Cogito 1 models can be downloaded or used via API with Cloud Provider Fireworks AI and AI together.
“We are currently still in the early stages of (our) scaling curves, which uses only a small portion of the calculations that are normally reserved for posting/continuous training on traditional large language models,” Kogito wrote in a blog post. “As we move forward, we are investigating complementary post-training approaches for self-improvement.”
San Francisco-based Deep Cogito was founded in June 2024, according to a California submission. The company’s LinkedIn page lists two co-founders, Drishan Arora and Dhruv Malhotra. Malhotra was previously the product manager for Google AI Lab DeepMind. Arora was a senior software engineer at Google.
According to Pitchbook, supporters Deep Kogito, including South Park Commons, are ambitiously aiming to build “general super intelligence.” The company’s founders understand the phrases that mean AI that can perform tasks better than most people, and “reveal completely new features that we haven’t yet imagined.”