# Recursive’s Flint: a free coding model for laptops, designed largely by AI

URL: https://ai-pulse.aglaganova.com/blog/recursives-flint-a-free-coding-model-for-laptops-designed-largely-by-ai
Published: 2026-10-21
Author: Roman Aglaganov, Full-stack developer
Category: Models
Region: Global
Tags: coding, open-weights, benchmarks

> Recursive, Richard Socher’s self-improving AI startup, released its first model. Flint is a 40-billion-parameter coding model under Apache 2.0, and its automated research system shaped much of how it was built.

![Illustration: a black spiral with dots along it, curling inwards to a red point at the centre](https://ai-pulse.aglaganova.com/api/media/file/flint-1.png)

Recursive, the startup founded by Salesforce’s former chief scientist Richard Socher, released its first model on 21 October. Flint is a coding model that runs on a laptop, is free to download and, the company says, was largely designed with the help of AI.

## The model

**Size:** a mixture-of-experts model with 40 billion parameters in total, of which about 2.5 billion are used for each token. That keeps it fast on ordinary hardware.

**Licence:** Apache 2.0, a true open-source licence, with weights on Hugging Face. It runs in common local tools such as llama.cpp, Ollama and LM Studio, so code never has to leave your machine.

**Speed:** Recursive says it reaches over 125 tokens a second on a MacBook Pro with an M4 Pro chip, and over 150 on Nvidia’s RTX Spark, up to twice as fast as comparable models.

**Scores:** by Recursive’s account, Flint leads mixture-of-experts models under 50 billion parameters on agentic coding tests including SWE-Bench Verified and Terminal-Bench 2.1. These are the company’s own claims.

## Built by the Eureka Machine

Recursive’s pitch is AI that improves AI. Its automated research system, the Eureka Machine, proposes ideas, runs experiments and picks next steps. For Flint, the company says it chose the final shape of the model, found a way to compress parts of it to 4 bits against common practice, nearly halving the training steps needed for that stage, and wrote custom code for Apple and Nvidia chips. A small team built the model from scratch in under four months. Recursive did not say how much human time the system saved.

## The limits

Recursive is upfront that Flint needs more guidance than the big paid coding assistants: it recommends breaking large tasks into smaller steps, and keeping prompts under 256,000 tokens. Wired’s Will Knight, who tried it on a 48 GB Mac, found it handled his requests reliably where smaller models stumbled.

## Why it matters

A good free coding model on a laptop means private code and no usage bills, which matters to anyone priced out of cloud assistants. The bigger story is how it was made. If automated research can produce a competitive model with a small team in four months, the pace of model releases is about to pick up again.

## Sources

- [Recursive: Introducing Recursive Flint](https://recursive.com/updates/introducing-recursive-flint)
- [TechCrunch: Recursive launches Flint, a laptop coding model it says was largely built by AI](https://techcrunch.com/2026/10/21/recursive-launches-flint-a-laptop-coding-model-it-says-was-largely-built-by-ai/)
- [Wired: Recursive’s Flint model](https://www.wired.com/story/recursive-self-improvement-flint-latest-model-scary-good/)