Deep Cogito has recently unveiled a series of open large language models (LLMs) that claim to surpass competitors, representing significant progress toward general superintelligence. Based in San Francisco, the company focuses on its mission to construct general superintelligence and has released preview versions of its LLMs in various sizes, ranging from 3B to 70B parameters.
According to Deep Cogito, each of these models has outperformed the best available open models in its class, including those from LLAMA, DeepSeek, and Qwen, across standard benchmarks. Notably, the 70B model even exceeds the performance of the newly launched Llama 4 with its 109B Mixture-of-Experts (MoE) architecture.
A key component behind this achievement is Deep Cogito’s innovative training method known as Iterated Distillation and Amplification (IDA). The company describes IDA as a scalable alignment strategy that fosters iterative self-improvement.
This method seeks to address the limitations posed by existing training paradigms, where the intelligence of a model often relies heavily on the capabilities of larger overseer models. IDA utilizes two core processes: amplification and distillation, which together form a feedback loop that can enhance model intelligence directly in proportion to computational resources, free from the constraints of overseer intelligence.
The capabilities of the newly released Cogito models, which are based on Llama and Qwen checkpoints, emphasize applications in coding, function calling, and agentic scenarios. Each model offers dual functionality, supporting both standard responses and self-reflective reasoning akin to advanced models like Claude 3.5.
However, Deep Cogito has indicated that they have not focused on optimizing for extended reasoning due to user preferences for quicker answers. Summary benchmark comparisons reveal that the Cogito models demonstrate substantial performance benefits over size-equivalent state-of-the-art models, particularly in reasoning tasks.
For example, the 70B Cogito model scores 91.73% on the MMLU benchmark in standard mode, outperforming its closest competitor, Llama 3.3. While acknowledging that benchmarks may not fully reflect practical applications, Deep Cogito is optimistic about the real-world capabilities of their models.
This release is described as a preview, with plans for further enhancements and larger models to be made available in the near future, all with open-source accessibility.