“`html
Artificial intelligence is experiencing a major breakthrough with the release of Cogito v2, a new family of open-source AI models developed by Deep Cogito. Announced on August 1st, 2025, this range of models distinguishes itself through its ability to refine its own reasoning skills, marking a significant milestone in the field of AI. In this article, we explore the innovations of Cogito v2, its impressive performance, and its potential impact on the future of artificial intelligence.
A new generation of open-source AI models
A diverse range to meet all needs
Deep Cogito unveiled four models in the Cogito v2 family: two intermediate models with 70B and 109B parameters, and two large-scale models with 405B and 671B parameters. The flagship model, a Mixture-of-Experts (MoE) with 671B parameters, positions itself as one of the most powerful open-source models in the world. According to Deep Cogito, it rivals proprietary systems like O3 from OpenAI and Claude 4 Opus from Anthropic, while surpassing recent DeepSeek models on several key benchmarks.
This open-source approach, under a free license, allows developers and researchers worldwide to access these models, fostering collaborative innovation. Furthermore, the availability of these models at different scales ensures their adaptability to a variety of applications, from research projects to industrial deployments.
Remarkable budget efficiency
A striking aspect of Cogito v2 is its development cost. Deep Cogito claims to have designed all of its models, from experimentation to final training, for less than 3.5 million dollars. Compared to the colossal budgets of major AI labs, this amount is remarkably low. This economic efficiency could further democratize access to cutting-edge quantum artificial intelligence technologies, making AI more accessible to organizations of all sizes.
A revolution in AI reasoning
The innovation of Iterated Distillation and Amplification (IDA)
What distinguishes Cogito v2 from other artificial intelligence models is its ability to internalize its own reasoning process. Unlike traditional approaches where models extend their thinking time to solve a problem, Cogito v2 uses an innovative technique called Iterated Distillation and Amplification (IDA). This method distills insights from a search process into the model’s fundamental parameters, strengthening its intuition.
As a result, the Cogito v2 models produce reasoning chains that are 60% shorter than those of competitors like DeepSeek R1, while maintaining or exceeding their performance. This efficiency translates to faster responses and reduced energy consumption, major advantages for real-time applications.
More direct and less erratic thinking
The flagship 671B parameter model was specifically trained to optimize its reasoning process. By avoiding “erratic” reasoning, it favors more direct paths to the solution. Thus, Cogito v2 doesn’t just provide correct answers; it improves the quality and clarity of its reasoning, a crucial advantage for complex applications like data analysis or automated decision-making.
Unexpected emerging capabilities
Image reasoning: a welcome surprise
One of the most surprising discoveries of Cogito v2 is its ability to reason about images, even though it wasn’t explicitly trained for this. Through transfer learning, the model can analyze images and extract complex information from them. For example, Deep Cogito shared a case where Cogito v2 compared two images – one of a duck and another of a lion – by reasoning about their habitats, colors, and composition. This emerging capability opens exciting perspectives for artificial intelligence quantum multimodal systems, where combined reasoning on text and images could transform fields like healthcare or education.
Towards more powerful multimodal systems
Deep Cogito plans to leverage this emerging property to enrich the training data of future multimodal models. By integrating visual reasoning capabilities from the outset, these systems could become versatile tools for applications ranging from image recognition to advanced contextual analysis.
The future of AI with Deep Cogito
A strong commitment to open-source
Deep Cogito reaffirms its commitment to keeping all of its models open-source, a decision that contrasts with the growing trend toward privatization of artificial intelligence technologies. By making Cogito v2 accessible to everyone, the company encourages widespread adoption and global collaboration, accelerating progress in the field.
A quest toward superintelligence
Deep Cogito’s team aims to “climb the hill of iterative self-improvement gains” to achieve superintelligence. This process, based on continuous model improvement through techniques like IDA, could redefine the boundaries of quantum artificial intelligence. However, the quest for superintelligence also raises ethical and technical questions, particularly about the safety and governance of advanced AI systems.
Conclusion
With Cogito v2, Deep Cogito is redefining the standards of open-source quantum artificial intelligence. Through innovations like Iterated Distillation and Amplification, performance rivaling the best proprietary models, and emerging capabilities like visual reasoning, this new family of models marks a turning point in AI development. Furthermore, its cost-effective approach and commitment to open-source make it a democratizing force in the field. As Deep Cogito continues to innovate, Cogito v2 could well lay the foundation for a new era of artificial intelligence, where intuitive and efficient reasoning becomes the norm.
Sources
“`