A new contender has quietly entered the AI race. Deep Cogito, a San Francisco-based startup, has emerged from stealth mode with a family of AI models called Cogito 1 that promise something rare in the open-source world. A toggle between fast, direct answers and deeper, step-by-step reasoning.
Unlike traditional language models that stick to a single operating mode, Deep Cogito’s hybrid architecture lets each model switch between “reasoning” and “non-reasoning” modes. This feature allows the models to adapt in real time. Delivering speedy responses to simple prompts while slowing down to think through complex ones. It’s a concept gaining momentum among AI research labs like Anthropic and OpenAI. Especially as demand rises for models that balance performance and efficiency.
All Cogito 1 models are hybrid by default, designed to dynamically shift their behavior based on the task. Deep Cogito says this makes them faster and smarter than many of the best open models currently available. Internally benchmarked results show that the company’s flagship model. Cogito 70B, outperforms DeepSeek’s R1 reasoning model on several math and language tasks. When reasoning is disabled, it even surpasses Meta’s Llama 4 Scout on LiveBench, a popular real-world AI benchmark.
What’s striking is the speed of development. Deep Cogito claims that the entire Cogito 1 family was built in just 75 days by a small team. The models weren’t trained entirely from scratch. Instead, the company fine-tuned existing open models like Meta’s Llama and Alibaba’s Qwen. Applying proprietary training techniques that enhanced their reasoning abilities and enabled the mode-switching feature.
Cogito 1 models span sizes from 3 billion to 70 billion parameters, and even larger models—up to a massive 671 billion—are on the horizon. While model size isn’t everything, more parameters often translate into more refined reasoning and better generalization. All models are openly accessible for download or use via API through platforms like Fireworks AI and Together AI.
Behind the ambitious project are co-founders Drishan Arora and Dhruv Malhotra. Arora previously worked as a senior engineer at Google, while Malhotra was a product manager at DeepMind. Where he helped develop Google’s generative search features. The company itself was incorporated in June 2024 and is backed by South Park Commons, according to PitchBook filings.
Despite being new to the scene, Deep Cogito is making no small claims about its long-term ambitions. The team says it’s working toward general superintelligence—AI that not only outperforms most humans on current tasks but also uncovers new abilities we haven’t even imagined yet. That vision is still a long way off, but their hybrid model architecture is an early glimpse into what a more flexible and adaptive generation of AI could look like.
The company is also exploring advanced post-training techniques to further refine its models, acknowledging that it has only used a fraction of the compute traditionally required for continued large-scale training. That leaves room for growth—and perhaps even more surprising breakthroughs.
In a field where most open-source efforts have struggled to match the performance of proprietary models, Deep Cogito’s fast-moving development cycle and technical edge could make it one of the most watched new players in AI.
