Meet Cogito 1: Fast, Thoughtful, and Open

A new player has entered the AI arena. Deep Cogito, emerging from stealth mode, has unveiled a family of open-source AI models designed with a twist — they can switch between reasoning and non-reasoning modes on demand.

This “hybrid architecture” approach lets the models decide when to think hard and when to answer fast, depending on the complexity of the question. It’s a strategy that addresses the trade-off between deep reasoning (ideal for complex tasks like math and logic) and speed/efficiency (preferred for quick, factual responses).

Meet Cogito 1: Fast, Thoughtful, and Open

All of Deep Cogito’s models, under the Cogito 1 brand, fall under this hybrid design. They reportedly outperform other open models of similar size, including Meta’s LLaMA line and DeepSeek’s offerings.

In a blog post, the company explained:

“Each model can answer directly [...] or self-reflect before answering (like reasoning models).”

Key specs and facts:

• Model sizes: 3B to 70B parameters, with 671B+ models on the way.

• Built from: Meta’s LLaMA and Alibaba’s Qwen models.

• Optimized using: Novel training techniques to enhance reasoning toggle capability.

The Cogito 70B model, in reasoning mode, beats DeepSeek’s R1 reasoning model in several math and language benchmarks. When reasoning is turned off, it still outperforms Meta’s LLaMA 4 Scout on LiveBench, a general-purpose benchmark.

Download and Use

Cogito 1 models are fully open and available via:

• Fireworks AI

• Together AI

Developers can either download the models or access them through APIs, making them highly accessible for experimentation and deployment.

Why This Matters

Reasoning-capable models are getting more attention — with companies like OpenAI (with o1) and Anthropic building agents that "think through" answers. But these models tend to be computationally heavy. Deep Cogito's hybrid approach is designed to give users the best of both worlds — speed when you need it, thoughtfulness when it counts.

The fact that Deep Cogito’s small team built these models in just 75 days adds intrigue — and pressure — to legacy labs already racing to dominate the open-source LLM space.

Bottom line:
Deep Cogito may be new, but its hybrid, open-source AI models could shake up how we think about balancing reasoning power with speed. If nothing else, they’re a bold bet on the future of flexible, developer-friendly AI.