New open source AI company Deep Cogito releases first models and they’re already topping the charts

Deep Cogito Emerges with Revolutionary AI Models

In a groundbreaking move, Deep Cogito, a cutting-edge AI research startup located in San Francisco, has officially unveiled its first line of open-source large language models (LLMs), named Cogito v1. These models, fine-tuned from Meta’s Llama 3.2, boast hybrid reasoning capabilities that allow them to respond swiftly or engage in introspective thinking—a feature reminiscent of OpenAI’s “o” series and DeepSeek R1.

Deep Cogito envisions pushing AI beyond conventional human oversight constraints by fostering iterative self-improvement within its models. Their ultimate goal? To develop superintelligence—AI that surpasses human capabilities across all fields. Yet, the company assures that all models will remain open-source.

Drishan Arora, CEO and co-founder of Deep Cogito, previously served as a Senior Software Engineer at Google, leading the development of LLMs for Google’s generative search product. He confidently stated on X that these models are among the strongest open models at their scale, outperforming competitors like LLaMA, DeepSeek, and Qwen.

The Model Lineup

The initial offering includes five base sizes—3 billion, 8 billion, 14 billion, 32 billion, and 70 billion parameters—and is already accessible on platforms such as Hugging Face, Ollama, and APIs via Fireworks and Together AI. These models operate under Llama licensing terms, allowing commercial use for up to 700 million monthly users before requiring a paid license from Meta.

Deep Cogito intends to roll out even larger models, potentially reaching 671 billion parameters, in the near future.

Training Approach: Iterated Distillation and Amplification (IDA)

Arora introduced IDA, a novel method distinct from traditional reinforcement learning from human feedback (RLHF) or teacher-model distillation. IDA focuses on allocating additional computational resources to generate superior solutions, subsequently embedding this enhanced reasoning into the model itself—a continuous feedback loop aimed at boosting capabilities. This approach mirrors Google AlphaGo’s self-play strategy adapted for natural language processing.

Benchmarks and Evaluations

Deep Cogito presented comprehensive evaluation results comparing Cogito models against open-source counterparts in areas such as general knowledge, mathematical reasoning, and multilingual tasks. Key findings include:

• Cogito 3B (Standard): Outperforms LLaMA 3.2 3B on MMLU by 6.7 percentage points (65.4% vs. 58.7%) and on Hellaswag by 18.8 points (81.1% vs. 62.3%).
• Cogito 3B (Reasoning Mode): Scores 72.6% on MMLU and 84.2% on ARC.
• Cogito 8B (Standard): Achieves 80.5% on MMLU, outscoring LLaMA 3.1 8B by 12.8 points.
• Cogito 8B (Reasoning Mode): Scores 83.1% on MMLU and 92.0% on ARC.
• Cogito 70B (Standard): Leads LLaMA 3.3 70B on MMLU by 6.4 points (91.7% vs. 85.3%) and surpasses LLaMA 4 Scout 109B on aggregate benchmarks (54.5% vs. 53.3%).

While Cogito models excel in reasoning mode, certain trade-offs exist, particularly in mathematical tasks.

Native Tool Calling

Deep Cogito also assessed its models’ native tool-calling performance, a crucial aspect for agent and API-integrated systems.

• Cogito 3B: Supports four tool-calling tasks and excels in simple and multiple tool calls.
• Cogito 8B: Demonstrates strong performance across all tool call types, outperforming LLaMA 3.1 8B significantly.

Future Plans

Looking forward, Deep Cogito plans to introduce larger models, including mixture-of-experts variants at 109B, 400B, and 671B parameters, alongside ongoing updates to existing checkpoints. The company views IDA as a sustainable pathway toward scalable self-improvement, reducing reliance on human or static teacher models.

Arora highlighted that real-world utility and adaptability are the ultimate measures of success, emphasizing that this is merely the start of a promising journey. Deep Cogito collaborates with renowned entities like Hugging Face, RunPod, Fireworks AI, Together AI, and Ollama, ensuring all models remain open-source and freely accessible.