The AI Stack
Sign in
L4 · Models & Training

Deep Cogito

Develops open-source hybrid reasoning LLMs using Iterated Distillation and Amplification.

Updated April 2026

Overview

Segment
Frontier Foundation Model Labs
Posture
Enterprise LLM

Product overview

Deep Cogito builds Cogito series LLMs (3B to 671B parameters) with toggleable reasoning modes, trained via IDA on Llama/Qwen bases for superior benchmark performance. These models outperform same-size open competitors like Llama, DeepSeek, Qwen in math, language, and tool-calling; used by developers via Hugging Face downloads, Ollama, and APIs from Fireworks AI and Together AI. Distinct from frontier labs via efficient IDA self-improvement (under $3.5M training), open-sourcing all models, and hybrid architecture avoiding always-on reasoning latency.

Revenue model

Open-source models (Llama license, commercial use up to 700M users); hosted inference APIs via providers like Together AI, Fireworks AI (e.g. Cogito 8B $0.20/M tokens; 70B $0.90/M); raised $13M seed from Benchmark.

Moat

Deep Cogito's competitive moat is its proprietary Iterated Distillation and Amplification (IDA) framework, which enables them to achieve frontier-level model performance at a fraction of industry costs (under $3.5 million versus hundreds of millions spent by competitors), creating a sustainable efficiency advantage that's difficult to replicate without the same novel training methodology and team expertise.