Built on RunPod: How Cogito Trained Models Toward ASI

RunPod has partnered with Deep Cogito to release Cogito v1, a range of open-source AI models from 3 billion to 70 billion parameters, outperforming leading alternatives like LLaMA and Qwen across most benchmarks. These models, trained in just 75 days using RunPod's infrastructure, employ a novel training method called Iterated Distillation and Amplification (IDA), which enhances the model's reasoning capabilities beyond traditional methods by refining its own intuition without human or larger model supervision. Cogito v1 supports dual modes: Direct Mode for fast, high-quality completions and Reasoning Mode for more nuanced responses, catering to diverse applications such as coding and tool use. Future developments include larger Mixture of Experts models up to 671 billion parameters. The models are available under an open license on platforms like Hugging Face and Ollama, exemplifying how post-training improvements can drive AI toward super-intelligence and new capabilities.