Recursive, an AI startup established last year, is living up to its name by developing artificial intelligence capable of self-improvement. On Wednesday, the company unveiled its inaugural model, Flint. This is a free coding model designed to run on laptops and was constructed with significant assistance from artificial intelligence.
Recursive claims that its automated research system, dubbed the Eureka Machine, contributed to nearly every stage of Flint's development. This capability enabled a small team to construct the model from the ground up within four months.
Flint represents an early public experiment in recursive self-improvement, a concept suggesting AI can evolve into superior versions that build upon previous iterations. This concept is highly debated within the industry, with some researchers expressing concern that AI automating its own progress could eventually slip beyond human control. For the present, Recursive's iteration remains constrained: an AI research assistant managed by humans that resulted in a compact model suitable for laptop use.
The company emerged from stealth mode in May, securing $650 million in funding at a $4.65 billion valuation. Backers included GV, Nvidia, and AMD Ventures. Recursive is led by CEO Richard Socher, the former chief scientist of Salesforce.
Since Flint operates on a user's personal computer rather than a corporate cloud server, developers can utilize it offline, maintain code confidentiality, and sidestep the usage fees associated with cloud AI services that charge based on text processing volume. The model is downloadable for free on Hugging Face under the Apache 2.0 license, an open-source framework permitting unrestricted use and modification.
Flint operates as a mixture-of-experts model, meaning it allocates only a specific subset of its resources to any given task. It utilizes just 2.5 billion of its 40 billion parameters at a time. This design keeps it lightweight enough for standard laptops and distinguishes it from Meta’s Muse Spark Open, an open model released recently that requires 816 billion parameters and a terabyte of memory.
Flint also demonstrates notable speed. In Recursive's internal tests, it generated approximately 131 tokens per second on a MacBook Pro equipped with Apple’s M4 Pro chip. This speed is roughly double that of similar models on the same machine, including Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning. On Nvidia's RTX Spark, Flint exceeded 150 tokens per second.
On the SWE-Bench Verified benchmark, which assesses an AI's ability to resolve real-world software bugs, Flint surpassed other models with fewer than 50 billion parameters. According to Recursive's average performance across six tests, only Alibaba’s Qwen3.8 27B scored higher than Flint.
The company acknowledged that Flint requires more guidance than larger, paid coding assistants and suggested breaking down extensive tasks into smaller components. Nevertheless, initial feedback has been favorable. Will Knight of *Wired* utilized Flint on a Mac with 48GB of memory to develop several small applications, including a to-do list and a research tool. He noted that smaller coding models often encounter errors or slowness compared to Claude or Codex, but Flint successfully executed every coding request he made.
The Eureka Machine operates by generating hypotheses, conducting experiments, and utilizing outcomes to determine the next steps. The company's launch video highlighted that Eureka was integral to nearly every aspect of Flint's creation, ranging from pre-training and post-training to inference.
Recursive stated that Eureka's contributions included a sparser model structure, low-bit training that reduced Flint's size by 71% with minimal quality loss, and custom kernels designed to accelerate underlying computations.
Although Recursive did not disclose the specific time saved by the Eureka Machine, it is not the only entity relying on AI for research. OpenAI reported last month that by mid-August, its AI agents were logging 3.1 workdays for every single day worked by a human researcher.
If Recursive's strategy proves successful, subsequent models should be capable of handling longer work cycles without human intervention. The company previously informed TechCrunch that its $410 million compute agreement with Amazon Web Services in June was "less about headcount and more about agent count."
The ultimate objective, as described by Socher to *Wired*, is for the Eureka Machine to achieve complete recursive self-improvement, with humans maintaining oversight to intervene if necessary. Recursive maintains that each new model enhances the Eureka Machine's capabilities as a researcher, thereby accelerating the development of the next generation.
The Editorial Staff at AIChief is a team of professional content writers with extensive experience in AI and marketing. Founded in 2025, AIChief has quickly grown into the largest free AI resource hub in the industry.