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

Deep Cogito has pulled in $43 million in fresh funding, a Series A led by TQ Ventures. Benchmark, Nexus Venture Partners, Atreides Management, South Park Commons and Zscaler also joined the round, lifting the startup’s total outside funding to more than $56 million.

The company is only from 2024, founded by former Google employees Drishan Arora and Dhruv Malrana, but it is aiming at a very specific prize: AI models that can autonomously improve themselves. That’s the pitch behind Cogito, its open-source large language model family, which first appeared in April 2025.

The latest model in that line, Cogito v2.1 671B, arrived last November. Deep Cogito said in a launch blog post that it was ahead of any other U.S. open model available at the time, and that it used fewer tokens than comparable reasoning models. In other words, it was trying to look both smart and cheap.

The company says its models are trained with an internal method called IDA, which increases the infrastructure available to a model during prompting, then uses the output to refine parameters. It also uses reinforcement learning, plus a process-supervision approach that grades the steps an LLM takes on the way to an answer instead of only checking the final response. Deep Cogito says that makes the models less likely to wander into pointless or wrong steps and helps cut hardware use.

There’s a business here too. Deep Cogito sells a platform that lets enterprises build custom AI models on their own data, and Zscaler is already a user. The new cash will go into that enterprise push, along with more research hires and new open-source releases. And the timing is hardly lonely: Recursive Superintelligence raised $650 million in May, while Ineffable Intelligence closed a $1.1 billion round in April.