Deep Cogito Raises $43M Series A to Build Self-Improving AI Models That Autonomously Advance

Deep Cogito raises $43M Series A for AI self-improvement research

Deep Cogito was founded in 2024 by Drishan Arora, who serves as chief executive, and Dhruv Malrana, chief product officer. The two co-founders had previously worked together at Google $GOOGL on AI Search -- Arora led Gemini post-training for AI Search, while Malrana led its product from inception, the company said. The startup's work centers on the post-training phase of AI development, the process that occurs after a model has been trained on large volumes of general data. Its research includes large-scale reinforcement learning and a technique called Iterated Distillation and Amplification, in which a model is given extra computation to generate answers it could not produce in one step, and those improved outputs are then folded back into the model's underlying parameters. Ultimately, the company aims to develop models capable of advancing their own intelligence independently, reaching a point where they are no longer constrained by the boundaries of human-generated training data.

Deep Cogito raises $43M to develop self-improving AI models

Artificial intelligence startup Deep Cogito Inc. today announced that it has raised $43 million in funding. TQ Ventures led the Series A round. It was joined by Benchmark, Nexus Venture Partners, Atreides Management, South Park Commons and Zscaler Inc., a publicly traded cybersecurity provider. The deal brings Deep Cogito's total outside funding to over $56 million. Deep Cogito was founded in 2024 by former Google LLC employees Drishan Arora and Dhruv Malrana. The company is looking to develop AI models that possess the ability to autonomously improve themselves. Its research efforts in that area have produced Cogito, a family of open-source large language models released in April 2025. The newest entry into the LLM series, Cogito v2.1 671B, debuted last November. The company stated in a launch blog post that the algorithm was "ahead of any other U.S. open model" available at the time. Cogito v2.1 671B also had lower token usage, a measure of LLM cost-efficiency, than comparable reasoning models. Deep Cogito trains its models using an internally developed technique dubbed IDA. When an LLM is given a prompt, IDA boosts the amount of infrastructure at the algorithm's disposal to increase its output quality. It then analyzes the output quality increase to identify ways of refining the model's parameters. Deep Cogito also trains its models with reinforcement learning. In a reinforcement learning run, an LLM is given sample tasks and receives feedback on the quality of its responses from a second neural network. That feedback helps the LLM hone its reasoning capabilities. Usually, reinforcement learning workflows only generate feedback about an LLM's finalized prompt responses. Deep Cogito uses a method called process supervision to grade the individual steps through which an LLM arrives at the answer. According to the company, that approach makes its models less likely to take unnecessary or erroneous steps, which lowers hardware usage. Deep Cogito monetizes its technology with a platform that enables enterprises to develop custom AI models using their internal data. Zscaler, one of the investors in today's round, is among the platform's users. Deep Cogito will use the new funding to grow its enterprise business. Additionally, the company plans to grow its research team and release new open-source models. Deep Cogito is one of several startups working to develop self-improving AI models. Another market player, Recursive Superintelligence Inc., raised $650 million in May from a consortium that included Nvidia Corp. In April, a newly launched competitor called Ineffable Intelligence Ltd. closed a $1.1 billion round at $5.1 billion valuation.