AI Startup Recursive Releases Flint, a Laptop-Ready Coding Model Built by AI

OPA WIRE, Oct 21 (OPA) — Recursive, a startup led by former Salesforce chief scientist Richard Socher, released Flint on Wednesday. The model is designed to operate offline on personal computers, allowing developers to maintain code privacy while avoiding cloud-based usage fees. It is available for download on Hugging Face under the Apache 2.0 license.

Flint utilizes a mixture-of-experts architecture, activating 2.5 billion of its 40 billion parameters at any given time. This efficiency allows it to outperform larger models on standard hardware. In internal tests on a MacBook Pro with an M4 Pro chip, Flint generated approximately 131 tokens per second, significantly faster than Google’s Gemma 4 or Nvidia’s Nemotron 3.5 Lightning. On Nvidia’s RTX Spark, the model exceeded 150 tokens per second.

The company credits its proprietary Eureka Machine for the rapid development of the model, which was built from scratch in under four months. According to Recursive, the system assisted in nearly every phase of the process, including pre-training, post-training, and the development of custom kernels. These optimizations reportedly reduced the model's size by 71% without substantial quality loss.

Recursive, which emerged from stealth in May with a $4.65 billion valuation, views Flint as a practical application of recursive self-improvement. While the current iteration functions as a human-supervised research assistant, the company aims to eventually automate more complex development cycles. Future iterations are expected to increase the Eureka Machine’s capacity to build subsequent models with decreasing levels of human intervention.