Overview
DeepCogito introduces the Cogito-v1 Preview series, a powerful suite of hybrid reasoning models trained with Iterated Distillation and Amplification (IDA). These models are designed to push the boundaries of open-weight LLMs through scalable alignment and self-improvement strategies, offering unmatched performance across coding, STEM, multilingual, and agentic use cases.
Each model in this series operates in both standard (direct answer) and reasoning (self-reflective) modes, significantly outperforming size-equivalent open models such as LLaMA, DeepSeek, and Qwen. The 70B variant notably surpasses the newly released LLaMA 4 109B MoE model in benchmarks.
Variants
Cogito-v1 Preview
| No | Variant | Branch | Cortex CLI command |
|---|---|---|---|
| 1 | Cogito-v1-Preview-LLaMA-3B | 3b | cortex run cognito-v1:3b |
| 2 | Cogito-v1-Preview-LLaMA-8B | 8b | cortex run cognito-v1:8b |
| 3 | Cogito-v1-Preview-Qwen-14B | 14b | cortex run cognito-v1:14b |
| 4 | Cogito-v1-Preview-Qwen-32B | 32b | cortex run cognito-v1:32b |
| 5 | Cogito-v1-Preview-LLaMA-70B | 70b | cortex run cognito-v1:70b |
Each branch contains a default quantized version:
- LLaMA-3B: q4-km
- LLaMA-8B: q4-km
- Qwen-14B: q4-km
- Qwen-32B: q4-km
- LLaMA-70B: q4-km
Use it with Jan (UI)
- Install Jan using Quickstart
- Use in Jan model Hub:
deepcogito/cognito-v1
Use it with Cortex (CLI)
- Install Cortex using Quickstart
- Run the model with command:
cortex run cognito-v1
Credits
- Author: DeepCogito
- Original License: Apache License 2.0
- Papers: Cognito v1 Preview