Recursive’s Flint Brings Fast Coding Models to Local Machines

Built by an automated research system.

@RichardSocher introduced Flint, @Recursive_SI’s first model: a fast local coding model built with the company’s automated research system, the Eureka Machine. He calls it best-in-class for fast local coding.

He claims it is up to 2x faster than comparable models, at 125+ tokens/sec on a MacBook Pro and 150+ on RTX Spark. It was built from scratch in under 4 months, he says. Open weights are available now on Hugging Face, and it can be tried for free on @opencode or @pidotdev. “Private, unmetered intelligence on your own machine,” he writes.

@MiaAI_lab wants a rematch.

@MiaAI_lab counts Flint at 40B params total, 2.5B active, and notes its claim of 150 tok/s on a DGX/RTX Spark. “Looks interesting enough to test,” she writes. Her question: could it beat “the legendary Qwen3.6-35B”? That model is still one of the best, if not the best, for its size, she says.