A new company in the AI world, Deep Cogito, has gained a spotlight and it’s already making waves. The San Francisco-based startup recently emerged from stealth mode, introducing a lineup of openly available AI models called Cogito 1. It comes with a unique twist, they can switch between “reasoning” and non-reasoning modes.
Reasoning models like OpenAI’s o1 are especially good at solving complex problems in subjects like math and physics because they “think” step by step. They basically double-check their own work before answering. That kind of self-reflection is super powerful but also requires more computing power and takes longer to respond.
That’s where hybrid models come in and it’s what Deep Cogito is betting big on. Instead of using heavy reasoning all the time, these hybrid models mix fast-response, non-reasoning layers with deeper, reasoning capabilities. The idea is to fly through simple questions but pause and reflect when things get complicated just like a person might.
The Tech Behind Cogito 1
Deep Cogito’s first family of models, Cogito 1, is a lineup of hybrid AI models ranging from 3 billion to 70 billion parameters. Bigger models are better at solving tough problems and Deep Cogito is planning to release models up to a staggering 671 billion parameters in the near future. This would put them in the same category as frontier labs like OpenAI and Anthropic.
What’s cool is that the Cogito 1 models aren’t built entirely from scratch. Instead, the team took open-source models from Meta (Llama) and Alibaba (Qwen) and enhanced them using novel training techniques. The model you can ask a quick question or tell it to “think it through” before answering.
In internal benchmarks, their model, Cogito 70B, outperformed Meta’s Llama 4 Scout and DeepSeek’s R1 on a range of language and math tasks even with reasoning turned off. When reasoning is turned on, it does even better.
You can already try them out or download them. They’re available via Fireworks AI and Together AI, two popular cloud providers for AI devs.
Related links you may find interesting
Built in 75 Days by a Tiny Team
Deep Cogito says it pulled all this off in just 75 days with a small team. The startup was founded in June 2024 by Drishan Arora (former senior software engineer at Google) and Dhruv Malhotra (former product manager at DeepMind). They’re backed by South Park Commons, a well-known founder community that has supported some pretty successful ventures in the past.
Dhruv’s time at DeepMind gave him direct experience working on generative search, a cutting-edge AI tech that Google is betting on for the future of Search. That insight probably gave Deep Cogito a head start in building models that are both fast and intelligent.
Aiming for AGI
Deep Cogito isn’t shy about its goals. The company wants to develop “general superintelligence” that’s a level of AI that outperforms humans across most tasks and even discovers new abilities that we haven’t imagined yet. While that sounds ambitious (and maybe a little sci-fi), the company seems to be taking a practical path by starting with tools developers can use today.
They admit they haven’t used anywhere near the amount of computing power that big players like OpenAI or Google throw at their models. But they’re already working on ways to enhance their models post-training, possibly through techniques like reinforcement learning, fine-tuning with human feedback, or self-improvement loops methods that could make the models smarter over time, even without massive compute budgets.
Why Does This Matter?
Right now, AI is at a crossroads. There’s a growing tension between open and closed models, with major companies such as OpenAI and Anthropic locking down their tech, while startups like Mistral, DeepSeek. Now Deep Cogito is pushing for openness and transparency.
Cogito’s hybrid approach, especially the ability to toggle reasoning, could offer a more efficient, cost-effective alternative to massive, power-hungry models. That’s a big deal for startups, educators, and researchers looking for smart models that don’t break the bank.
If Deep Cogito keeps up this momentum, it could become a serious contender in the open-source AI scene and might just help shape the future of AI development.
