Cogito
Open Models with Switchable Reasoning Mode
5 followers
Open Models with Switchable Reasoning Mode
5 followers
Cogito from Deep Cogito are the open models (3B-70B) with switchable reasoning. Trained via Iterated Distillation & Amplification (IDA), outperforming class benchmarks. Llama Community License.
Flowtica Scribe
Hi everyone!
Sharing Cogito v1 Preview, a new family of language models (3B to 70B parameters) from the recently emerged startup Deep Cogito. They're taking an interesting approach with these open models.
The key idea here is Hybrid Reasoning:
Direct Mode: Get fast answers like a standard LLM.
Reasoning Mode: Activate a "deep thinking subroutine" (via system prompt) where the model self-reflects before answering, similar to models like DeepSeek-R1. You choose the mode based on your needs!
They developed these using a technique called Iterated Distillation and Amplification (IDA), which aims for scalable self-improvement potentially beyond human data limits.
Performance-wise, Deep Cogito reports each Cogito model outperforms other open models in its size class (LLaMA, DeepSeek, Qwen), with the 70B version reportedly beating the Llama 4 109B MoE model and rivaling GPT-3.5/4o-mini on benchmarks.
These are open models using the Llama Community License, so check the terms for your specific use case.