Deep Cogito has entered the sector of artificial intelligence models (AI) with the presentation of its open source models V1which are based on hybrid models capable of alternating between Reasoning capabilities and instantaneous responses, surpassing the equivalent models of Meta or Deepseek in performance, which seeks to move towards general superintelligence.
The technology was founded in June last year based in San Francisco (California, United States) and intends Develop a “general superintelligence” Able to perform any type of task, “not only to match human capacities, but also to discover completely new abilities”, through advanced reasoning and “iterative self -consumption.”
In this sense, I think of V1 It is presented as a series of large models (LLM) open source, that are based on a hybrid functioningthat is, that they allow to use their capabilities to offer direct answers to simple applications, as well as for reflect Faced with more complex issues, in order to use their resources more optimally.
As the company has pointed out in a statement in its blog, Cogito V1 has been presented in the sizes 3B, 8B, 14B, 32B y 70B, All of them developed by “a Small team “in approximately 75 days.
Although these models are Based on the open models of Meta and Qwen de Alibaba, Deep Cogito has assured that SuityThey were “the best open models of the same size”, which includes the flame models developed by goal, as well as Depseek models.
Specifically, the company has stressed that the 70b Cogito V1 model exceeds the recently announced call 4 109b moe in the General Livebench purpose test. Following this line, according to the shared evaluation results, the largest version It also exceeds the Deepseek R1 reasoning model in most mathematics and language tasks.
To achieve these capabilities, Deep Cogito has clarified that their models have been trained using Iterated distillation and amplification methods (IDA), “A scalable and efficient alignment strategy for general superintelligence through iterative self -managing.”
As explained, when using distillation and amplification, they use more computational resources so that the model reaches a better solution and, after that, they manage to reduce the thought process to the model’s own parameters.
“In a half that the LLM improves its intelligence, the thought process itself becomes more powerful “, They have sentenced, while determining that a positive feedback cycle is created in which the model of the model is increasingly determined by computational resources and the effectiveness of the amplification and distillation process, instead of “by the limitations of the original supervisor.”
With all this, these models Cogito V1 They are already available for download through Huggingface, Ollama, or through API in Fireworks AI or Together AI. Likewise, Deep Cogito has indicated that, in the face of the coming months, they will launch larger models, including the sizes of 109b, 400b, 671b.