Deep Cogito Emerges from Stealth with Hybrid AI ‘Reasoning’ Models

Read Time: 2 minutes

AI startup Deep Cogito has emerged from stealth with Cogito 1, a family of hybrid models that can toggle between fast direct responses and in-depth reasoning. Built on Meta’s Llama and Alibaba’s Qwen, these models range from 3B to 70B parameters and outperform top open models in several benchmarks.

A new player in the AI landscape, Deep Cogito, has officially stepped out of stealth mode with an ambitious line of hybrid AI models that can toggle between standard and reasoning-driven modes—offering a new paradigm for how machines solve complex problems.

What Makes Cogito Models Unique?

The new Cogito 1 family, developed in just 75 days, includes models ranging from 3B to 70B parameters, with a roadmap extending to 671B parameters in the near future. What sets them apart is a “reasoning switch”: models can operate in either a fast, direct-response mode or a slower, step-by-step reasoning mode—ideal for complex domains like math, physics, and logic.

“Each model can answer directly […] or self-reflect before answering,” says Deep Cogito.

These models were not built entirely from scratch. Instead, the team enhanced open-source architectures like Meta’s Llama and Alibaba’s Qwen, using proprietary training techniques to unlock new performance capabilities.

How Do They Perform?

Cogito 70B with reasoning outperforms DeepSeek’s R1 model on key math and language benchmarks.
With reasoning disabled, it still outperforms Meta’s Llama 4 Scout on LiveBench, a general-purpose AI evaluation.
All models are now available via Fireworks AI and Together AI APIs.

The Hybrid Model Advantage

Hybrid models are gaining momentum as they balance speed and depth:

Fast responses for simple queries.
Thoughtful, deliberate reasoning for complex problems.
Reduced computational costs compared to always-on reasoning models.

This approach mirrors strategies at labs like Anthropic, which are exploring similar architectures for enterprise-grade AI solutions.

The Bigger Picture: General Superintelligence

Founded in June 2024 by Drishan Arora and Dhruv Malhotra (former Google DeepMind and Google AI engineers), Deep Cogito is backed by South Park Commons and has its sights set high:

“Our mission is to build general superintelligence—AI that can outperform most humans and unlock capabilities we’ve yet to imagine.”

While still early in its development arc, Deep Cogito claims it has used only a fraction of the typical compute required for major LLMs—hinting at big scalability potential.