Deep Cogito Raises $43 Million in Series A Funding

Deep Cogito announced the completion of a $43 million Series A funding round on August 26. The company develops post-training technologies for artificial intelligence models. TQ Ventures led the investment round. Benchmark, Nexus Venture Partners, Atreides Management, South Park Commons, and Zscaler also participated in the financing. The total amount raised by the company has now exceeded $56 million. The outlet BRIDGE reported the story. Deep Cogito focuses on enhancing the reasoning and task-specific capabilities of pre-trained models through additional training. The company researches reinforcement learning and mechanisms that allow models to improve their own abilities.

One key area of research is Iterated Distillation and Amplification, a process where models use extra computation to refine answers and then learn from those improvements. This iterative process aims to incorporate capabilities gained through additional computation directly into the model itself. The long-term goal is to surpass the limitations of human-created training data. The company applies this post-training approach to both its open-weight Cogito series and proprietary models trained on enterprise data. In November of last year, Deep Cogito released Cogito v2.1, which was built by applying its own post-training techniques to a DeepSeek foundation model. The company has made the trained weights of this model publicly available.

Zscaler, which participated in the funding round, previously collaborated with Deep Cogito as a customer. A Zscaler executive stated that general-purpose models lacked the necessary specialized expertise. The companies shared product details and key metrics to train the model on specific capabilities. Deep Cogito plans to use the new funds to expand its research and development staff. The company will also enhance the infrastructure used for training models. It intends to develop the next version of the Cogito model while increasing its work with enterprises seeking specialized models based on their own data.