A new AI player has entered the open-source race—and it’s bringing a fresh twist. San Francisco-based startup Deep Cogito has just unveiled its Cogito 1 family of AI models, featuring a unique toggleable reasoning mode that allows users to switch between fast, direct responses and more thoughtful, step-by-step problem solving.

Dubbed “hybrid AI models,” Cogito 1 models aim to combine the speed of conventional language models with the rigor of reasoning-first systems like OpenAI’s o1 or Anthropic’s Claude. This breakthrough has the potential to dramatically shift how developers and businesses use generative AI across domains from math and science to customer support and content generation.

“Each model can answer directly or self-reflect before answering, like reasoning models,” said Deep Cogito in a launch blog post.

What Are Hybrid AI Models, and Why Do They Matter?

Reasoning models like OpenAI’s o1 are praised for their ability to break down complex tasks—especially in technical subjects like math and physics—but that precision often comes with a cost: higher compute requirements and slower response times.

Hybrid models, like those from Deep Cogito, optimize for both performance and flexibility. They can:

• Rapidly handle simple queries in non-reasoning mode
• Switch to deeper, self-reflective logic for complex tasks

This structure enables smarter resource allocation without compromising quality—a holy grail for both developers and AI researchers.

Inside the Cogito 1 Model Family

The Cogito 1 models range from 3B to 70B parameters, with future releases scaling up to 671B parameters in the pipeline. Larger models tend to handle more complex reasoning and show stronger benchmark results.

The models weren’t built entirely from scratch. Instead, Deep Cogito customized and fine-tuned Meta’s Llama and Alibaba’s Qwen models using novel training techniques that enable this reasoning toggle capability.

According to internal benchmarks:

• Cogito 70B (with reasoning) beats DeepSeek’s R1 reasoning model in math and language evaluations.
• Cogito 70B (without reasoning) outperforms Meta’s Llama 4 Scout on LiveBench, a real-time AI performance benchmark.

All models are freely downloadable or accessible via Fireworks AI and Together AI APIs, making them an open, developer-friendly alternative in the evolving foundation model ecosystem.

Small Team, Big Ambitions

Incredibly, Deep Cogito says its team developed the Cogito 1 models in just 75 days—a pace rarely seen in the AI world.

“We’re still in the early stages of our scaling curve,” the company wrote. “We’ve used only a fraction of the compute typically needed for traditional post-training.”

The startup plans to expand by incorporating complementary post-training techniques for self-improvement—a signal that Cogito 1’s capabilities could grow even more rapidly.

Who’s Behind Deep Cogito?

Founded in June 2024, Deep Cogito is backed by South Park Commons and is led by two ex-Googlers:

• Dhruv Malhotra, formerly a product manager at DeepMind, where he worked on generative search
• Drishan Arora, previously a senior software engineer at Google

Their vision? Building “general superintelligence”—AI that not only outperforms most humans in many domains but may also uncover entirely new capabilities yet to be imagined.

What’s Next for Cogito and the Open AI Landscape?

The debut of Deep Cogito signals a new chapter in the race toward open, capable, and customizable AI models. As more startups prioritize hybrid reasoning, open distribution, and scalable architectures, the competitive landscape is shifting away from closed AI giants.

With a toggleable reasoning switch, developer-ready APIs, and an aggressive roadmap to reach models with hundreds of billions of parameters, Deep Cogito could become a serious challenger to Meta, DeepSeek, and even the front-runners in open-source AI innovation.

As the field barrels toward AGI, startups like Deep Cogito are asking a key question: What if your model could think more deeply—only when it needs to?

Get the Latest AI News on AI Content Minds Blog