Deep Cogito Inc. has officially launched its series of language models, claiming their performance surpasses that of comparably sized open-source alternatives. Founded in June 2024 by former Google LLC employees Drishan Arora and Dhruv Malhotra, the startup aims to carve out a niche in the rapidly evolving field of artificial intelligence.
On April 8, 2025, Deep Cogito unveiled its Cogito v1 series of open-source language models, which are available in five different sizes, ranging from 3 billion to 70 billion parameters. These models are based on the widely recognized open-source Llama and Qwen language model families, developed by Meta Platforms Inc. and Alibaba Group Holding Ltd., respectively.
The company has adopted a hybrid architecture for its models, allowing them to respond to user prompts either instantly or through more elaborate reasoning processes, depending on user preferences. This flexibility is designed to enhance user experience and cater to a variety of use cases.
One of the standout features of Deep Cogito's models is their innovative training method, referred to as IDA. This technique bears similarities to the widely used distillation process, which is employed to create hardware-efficient language models. In traditional distillation, a collection of prompts is sent to a hardware-intensive large language model (LLM), which generates answers. These answers are then utilized to improve a more efficient model, allowing it to respond to similar queries using less computational power.
However, Deep Cogito's IDA method takes a different approach. Instead of using the answers to enhance a separate model, it focuses on improving the original LLM that generated the responses. The IDA workflow consists of two primary steps. First, the LLM generates an answer to a prompt using reasoning methods that require more time to process data. Once the response is ready, the LLM distills the enhanced intelligence back into its parameters, effectively internalizing the improved capability.
Deep Cogito elaborated on the iterative nature of this process, explaining that each cycle builds upon the progress made in the previous one, creating a positive feedback loop that continually enhances the model's performance.
In an internal evaluation, Deep Cogito compared its most advanced model, which features 70 billion parameters, with Meta's Llama 3.3, which has the same number of parameters. The results indicated that Deep Cogito's model outperformed Llama 3.3 across all seven benchmarks utilized during the assessment. This claim of superiority extends to its smaller models as well, which are also said to outperform similarly sized open-source alternatives.
Deep Cogito's offerings include models with 3 billion, 8 billion, 14 billion, and 32 billion parameters, showcasing a diverse range of capabilities suited for various applications. The startup has ambitious plans for the future, with intentions to release new models in the coming weeks that will feature an impressive range of parameters from 109 billion to 671 billion.
As the landscape of artificial intelligence continues to evolve, the launch of Deep Cogito's models could signify a significant step forward in the competition among language models. With their innovative approach to training and their commitment to providing high-quality outputs, Deep Cogito is positioning itself as a formidable player in the AI industry.
In conclusion, Deep Cogito Inc. is making waves in the AI sector with its new language models, which promise to deliver superior performance while catering to a variety of user needs. The company's focus on continuous improvement through its IDA training method and its commitment to open-source principles may well set the stage for future advancements in language model technology.
