[CLSTR]
< Back to all clusters
[TECHNOLOGY] · 3 sources

started · updated

Recursive releases Flint, an AI coding model developed via automated research

Recursive Flint AI Eureka Machine mixture-of-experts coding models

AI startup Recursive has released Flint, its first language model designed for agentic coding on local hardware. The model was developed in under four months using the Eureka Machine, an automated research system that proposes, tests, and plans experiments. This process represents a public test of recursive self-improvement, where an AI assists in creating more performant models under human supervision.

Flint is a mixture-of-experts (MoE) model with 40 billion total parameters, though only 2.5 billion are active per token. This architecture allows it to run efficiently on laptops. In internal tests, the model generated approximately 131 tokens per second on a MacBook Pro with an M4 Pro chip and over 150 tokens per second on an Nvidia RTX Spark.

According to Recursive, Flint leads several coding benchmarks for MoE models under 50 billion parameters, including SWE-Bench Verified, Terminal-Bench 2.1, and SWE-Bench Pro v2. While some larger dense models like Qwen 3.8 may score higher in specific tests, Flint is designed to compete with 25 to 40 billion parameter dense models while operating significantly faster. The model is available for free on Hugging Face under an Apache 2.0 license.

Entities

Eureka Machine · Flint · Hugging Face · Recursive · Richard Socher

Claims

What the coverage asserts, and how many sources carry each claim.

Sources

Recursive lancia Flint: modello di coding per laptop [www.forux.it]
Recursive zeigt mit Flint erstes Ergebnis seines KI-gestützten Forschungsprozesses [the-decoder.de]
Recursive shows first result of its AI-driven research process with Flint [the-decoder.com]