Recursive Startup Unveils 40-Billion Parameter Model Built via Automated AI Pipeline

The drive toward fully autonomous artificial intelligence development has advanced a step further with the release of Flint, a new open-weight coding model designed to assist software engineers.

Flint is the debut offering from Recursive, a company founded by Richard Socher, the former chief scientist at Salesforce, and Jeff Clune, a professor at the University of British Columbia who studies methods for building AI systems capable of indefinite learning. Recursive's stated mission is to utilize recursive self-improvement (RSI) to advance both artificial intelligence and other scientific fields. The model is available for download and use on consumer hardware, specifically systems equipped with 48 GB of RAM.

The development of Flint relied heavily on an internal system Recursive calls the Eureka Machine, which was employed to automate the construction of the model. Socher indicated that nearly every component of the model pipeline was executed and refined by AI. This automation enabled a group comprising only a handful of researchers to produce a competitive model within a few months. According to the company, the Eureka Machine generated specific technical innovations, including new methods for training the base model, novel model architectures, and code optimized for particular silicon chips. Recursive has released benchmark scores suggesting Flint outperforms other models of similar size and architecture when handling coding tasks.

The concept of recursive self-improvement involves creating a loop in which AI identifies necessary improvements and initiates subsequent rounds of refinement. While this technique is currently constrained by human intervention, the process could theoretically continue indefinitely if AI contributes to every aspect of model development. However, this trajectory is not viewed positively by all experts. Some warn that allowing AI to improve itself in this manner could lead to disaster, arguing that an AI capable of indefinite improvement might escape human control or determine that humans obstruct its goals. These concerns gained broader attention in September when a researcher departed Anthropic and cautioned that advances in AI, including recursive self-improvement, could result in human extinction within a few years, a sentiment echoed by other experts. The idea has become controversial enough that several bills in Congress propose banning it.

Recursive characterizes such fears as more science fiction than reality and views legislation prohibiting RSI as an unnecessary hindrance to progress. Socher stated that Flint was constructed using a system incapable of escaping human control. While the Eureka Machine is intended to eventually achieve full recursive self-improvement, Socher emphasized that this process would remain under human supervision, allowing individuals to stop the operation if it fails. “We hope that this shows people that indeed we have all the steps in place” to build such a system, he said.

Recursive has secured $650 million in funding at a $4.65 billion valuation from investors including Google Ventures, Greycroft, Nvidia, and AMD Ventures. In testing, Flint, which has 40 billion parameters, demonstrated capabilities exceeding those of other models of its size. While larger models possess several trillion parameters and require industrial server racks, Flint operates on consumer hardware and has been reported to complete coding requests without the errors or slowness sometimes seen in comparable tools.

It remains unclear how much progress other U.S. labs have made regarding recursive self-improvement, though published research suggests fully recursive systems may be feasible. A team from Microsoft, Carnegie Mellon University, and the University of Virginia recently published work demonstrating how an AI coding agent can recursively improve itself by modifying the software harness that transforms reasoning into useful action and by modifying its own model weights. Socher suggested the approach used to build Flint could democratize access to advanced artificial intelligence. “It's the only way this very small team can produce a model that's actually better than any other model in its class,” he said. “This will hopefully show people that [recursive self-improvement] is actually quite nice and will benefit a lot of people.”