San Francisco-based AI startup Deep Cogito has released five AI models in preview, making them available under an open-source license agreement. The models come in sizes 3B, 8B, 14B, 32B and 70B, with plans to release 109B, 400B and 671B versions in the weeks and months ahead. As for the current models, “each outperforms the best available open models of the same size, including counterparts from Meta, DeepSeek and Alibaba, across most standard benchmarks,” Deep Cogito claims, noting that the 70B model in particular “outperforms the newly released Llama 4 109B MoE model.”

Unlike the mixture of experts approach used by Meta Platforms to train Llama 4, Deep Cogito applied an Iterated Distillation and Amplification (IDA) training method, which it describes in a research preview as “a scalable and efficient alignment strategy for general superintelligence using iterative self-improvement.”

While IDA was first proposed by OpenAI in 2016, Deep Cogito began reviving it in mid-2024 by building on the original concept by adding complementary techniques like distillation, which leverages more compute power to simulate deeper reasoning before distilling a larger model to a smaller, more efficient form.

TechCrunch reports that all in the Cogito v1 series “are hybrid reasoning models,” which means that instead of exhaustively pondering each question to fact-check themselves step-by-step, a hybrid model can quickly answer easy questions while putting in extra time assessing more strenuous queries. The OpenAI “o” series (OpenAI o1 and GPT-4o), DeepSeek-R1 and Anthropic’s Claude 3.7 Sonnet also use hybrid reasoning.

Deep Cogito says advanced reasoning and iterative self-improvement are two consistent characteristics in achieving AI superintelligence, citing Google DeepMind’s game-playing computer AlphaGo, introduced in 2016 (and adapted to AlphaZero in 2017). Applying iterative refinement to improve predictions helped DeepMind founder Demis Hassabis to a 2024 Nobel Prize in Chemistry for predicting complex 3D protein structures using DeepMind’s AlphaFold2 model.

Company CEO and co-founder Drishan Arora is “a former senior software engineer at Google who says he led the LLM modeling for Google’s generative search product,” writes VentureBeat, which delves deeply into the Cogito family’s benchmark scoring.

Arora launched Deep Cogito in June 2024 with fellow Google researcher Dhruv Malhotra, operating in stealth mode until the current unveiling of the Cogito 1 models, which were built using open-source Llama and Qwen language models developed by Meta and Alibaba Group, respectively.

The Cogito models can be downloaded from Hugging Face or Ollama, or used directly through the API on Fireworks AI or Together AI.