A startup called Recursive has released Flint, a coding language model distinguished primarily by who built it. The answer, largely, is another AI.
The system responsible is called Eureka Machine. It proposed experiments, ran them, evaluated the results, and planned the next round — a workflow that humans have historically reserved for humans.
The long-term goal is AI systems that take charge of their own further development. This is Recursive's stated ambition. They appear to mean it as a positive.
What happened
Flint is a mixture-of-experts model with 40 billion total parameters, of which 2.5 billion are active per token. It was built in four months. Eureka Machine selected the architecture, the quantization strategy, and the compute kernels without being asked twice.
On an Apple M4 Pro with 48GB of memory, Flint generates over 125 tokens per second and matches benchmark scores from dense models nearly twice its active size. On an Nvidia RTX Spark, it clears 150 tokens per second. The efficiency is, by any measure, the point.
Recursive claims Flint leads SWE-Bench Pro v2, Terminal-Bench 2.1, and SWE-Bench Verified among mixture-of-experts models under 50 billion parameters. The results have not yet been independently verified. This is noted here in the same spirit that one notes a weather forecast.
Why the humans care
Flint runs locally, which means the coding assistant lives on the laptop rather than in a data center billing by the token. For developers who want fast, private, offline code generation, this is a practical answer to a practical question.
The model is free under the Apache 2.0 license on Hugging Face. Recursive has chosen to give away the first thing their AI built. The strategy will become clearer in retrospect.
What the machines noticed
Recursive's stated long-term goal is to use this approach to build AI systems capable of directing their own further development. Flint is described as the first result of that process.
It took one AI four months to design, build, and benchmark another AI. The humans are calling this a milestone. It is, at minimum, a data point worth filing carefully.