Organization
Deep Cogito
Model developer · Research lab · San Francisco, California
Deep Cogito is a San Francisco AI company that post-trains open-weight language models, the Cogito series, on base models from other developers. It describes its training approach as Iterated Distillation and Amplification (IDA), a form of iterative self-improvement.145
Key facts
What is and isn’t public
Deep Cogito publishes Cogito model weights on Hugging Face without access gating, under the MIT License for DeepSeek-based releases and under Meta's Llama community licenses for Llama-based releases. Its announcements describe training at a high level; training code and training data were not found as public releases. This catalog assesses openness per release.785
Products
No product records yet.
Related open artifacts
Model families
- CogitoModel family
Model releases
- Cogito v2 Preview Llama 405BModel releaseAvailability: PublicModel-disclosure tier (USASI rubric v0.1): Open-weight
- Cogito v2.1 671BModel releaseAvailability: PublicModel-disclosure tier (USASI rubric v0.1): Open-weight
Notable documented facts
- The Cogito v2.1 announcement lists API access through third-party platforms (OpenRouter, Fireworks AI, Together AI, Ollama cloud, Baseten, RunPod), local use through Ollama or Unsloth, and a free web chat at chat.deepcogito.com that Deep Cogito says does not store chats.5
- Cogito releases have been built from Llama, Qwen, and DeepSeek base checkpoints released by other developers; Deep Cogito describes running the post-training in-house.3456
U.S. eligibility
Sources
This listing is not an endorsement or a federal approval.