Deep Cogito Unveils Advanced Open LLMs with IDA Methodology

Deep Cogito has released a series of open large language models (LLMs) that outperform existing models of similar sizes, using a novel training method called Iterated Distillation and Amplification (IDA). The models, available in sizes ranging from 3 billion to 70 billion parameters, are said to surpass counterparts from LLaMA, DeepSeek, and Qwen across most standard benchmarks.

Central to this release is the IDA methodology, which Deep Cogito describes as a scalable and efficient alignment strategy for general superintelligence. This technique involves iterative self-improvement through two key steps: amplification, which uses more computation to derive better solutions, and distillation, which internalizes these capabilities back into the model's parameters.

The 70B model from Deep Cogito notably outperforms the recently released Llama 4 109B Mixture-of-Experts (MoE) model. The company claims that IDA allows for a positive feedback loop where model intelligence scales more directly with computational resources, rather than being limited by the intelligence of larger overseer models or human curators.

The new models are optimized for coding, function calling, and agentic use cases, and feature dual functionality for direct answering or self-reflection before answering. Deep Cogito plans to release improved checkpoints and larger MoE models in the coming months. The models are available under an open license on Hugging Face, as announced by Victor M on X.