# Deep Cogito

> Deep Cogito is a San Francisco AI research company building open-source large language models with hybrid reasoning - the same model can answer instantly or stop to reason before responding. Founded in 2024 by former Google engineer Drishan Arora, the company trains its Cogito models with a method called Iterated Distillation and Amplification (IDA), which lets a model improve by learning from its own reasoning rather than from human labels. Its stated goal is general superintelligence, and its flagship 671B open model competes with the strongest open systems from DeepSeek and Meta.

- **Founded:** 2024
- **Headquarters:** San Francisco, California, United States
- **Founders:** Drishan Arora (Co-founder & CEO), Dhruv Malhotra (Co-founder)
- **Team size:** Small - roughly 10-15 people (reported ~11 employees); founder describes a small research and engineering team
- **Products:** Cogito v2 (open models), Cogito v1, Cogito v2.1 671B, Cogito Chat, IDA (training methodology)
- **Notable:** Emerged from stealth in April 2025 with open models that quickly topped open-model leaderboards, Trained its initial model family in roughly 75 days with a small team, Released a 671B MoE open model matching or exceeding DeepSeek v3/R1 and approaching o3 and Claude Opus

## Products & services

- **Cogito v2 (open models)** — A family of open-source hybrid reasoning LLMs spanning 70B, 109B (MoE), 405B, and 671B (MoE) parameters. Each model can answer directly or self-reflect before answering, toggling between standard and reasoning modes.
- **Cogito v1** — The company's first release (April 2025), fine-tuned from Meta's Llama and Alibaba's Qwen models, ranging from 3B to 70B parameters, with hybrid reasoning built in.
- **Cogito v2.1 671B** — A frontier open model forked from the open-licensed DeepSeek base, positioned among the strongest open models and approaching closed frontier systems like o3 and Claude Opus.
- **Cogito Chat** — A public chat interface at chat.deepcogito.com for trying the models directly.
- **IDA (training methodology)** — Iterated Distillation and Amplification - the training approach behind the models, in which a model amplifies its own reasoning and distills it back into itself to improve without human-labeled data.

## Achievements

- Emerged from stealth in April 2025 with open models that quickly topped open-model leaderboards
- Trained its initial model family in roughly 75 days with a small team
- Released a 671B MoE open model matching or exceeding DeepSeek v3/R1 and approaching o3 and Claude Opus
- Demonstrated ~60% shorter reasoning chains than DeepSeek R1 on reasoning tasks
- Put a long-theorized method (IDA) into practical, shipped form
- Raised a $13M seed round led by Benchmark

## Latest updates

- **2025-04** — Deep Cogito emerges from stealth with Cogito v1 hybrid reasoning models (3B-70B).
- **2025-08** — Closed a $13M seed round led by Benchmark, with South Park Commons participating.
- **2025-08** — Released the Cogito v2 family (70B, 109B MoE, 405B, 671B MoE) with self-improving 'intuition'.
- **2025-11** — Released Cogito v2.1 671B, forked from the open-licensed DeepSeek base model.

## Links

- Website: https://deepcogito.com
- LinkedIn: https://www.linkedin.com/company/deep-cogito/
- Twitter/X: https://x.com/deepcogito

---

Profile page: https://yespress.io/deep-cogito
Published by YesPress — https://yespress.io
Last updated: 2026-07-06