SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsAI developer platformsCogito

Cogito

Cogito represents a breakthrough in large language models (LLMs) through its Iterated Distillation and Amplification (IDA) training approach—a self-improving AI methodology that outperforms traditional models. The open-source release includes 3B to 70B parameter models, each surpassing counterparts

Open website

About this tool

Purpose

Cogito is a series of open-source large language models (LLMs) designed to explore the path to general superintelligence through Iterative Distillation and Amplification (IDA), a scalable and efficient alignment strategy. IDA leverages iterative self-improvement, allowing model capabilities to no longer be limited by the intelligence ceiling of external supervisors.

Key Capabilities

  • Provides models at five parameter scales: 3B, 8B, 14B, 32B, and 70B, all trained based on Llama / Qwen pretrained checkpoints.
  • Each model outperforms the best open-source models of the same scale on most standard benchmarks, including the corresponding versions of LLaMA, DeepSeek, and Qwen. The 70B model also outperforms the newly released Llama 4 109B MoE model.
  • Each model supports two modes: direct answering (standard LLM mode) or self-reflection before answering (similar to reasoning model mode).
  • Optimized for coding, function calling, and agent use cases.
  • Not optimized for extremely long reasoning chains, in order to balance users' need for response speed in practical tasks and to facilitate distillation back into model parameters in the next step.
  • Plans to release larger models (109B, 400B, 671B MoE) and improved checkpoints at each size, and all models will be open-sourced.

How to Use

  • Models can be downloaded on Hugging Face or Ollama.
  • Can be used directly through the Fireworks AI or Together AI APIs.
  • Each model can choose between answering directly or thinking before answering, similar to Claude 3.7's mode selection.

Applicable Scenarios

  • Coding tasks
  • Function calling (tool calling)
  • Agentic use cases
  • Scenarios requiring the fast responses of a standard LLM
  • Reasoning scenarios requiring the model to self-reflect before answering

Limitations

  • Currently a preview version (Cogito v1 Preview), still in an early stage, and has used only a small fraction of the compute typically consumed by conventional LLM post-training/continued training.
  • Specific information such as pricing, costs, and complete usage steps is not yet available.
  • Benchmark scores are only loosely correlated with how well a model meets user needs; the best evaluation should be close to users' actual needs.
  • Not optimized for extremely long reasoning chains.
  • Larger models (109B, 400B, 671B) and improved checkpoints have not yet been released.

Related tools

Browse category