Recursive unveils Flint, an AI coding model for laptops

Key Takeaways

1Recursive releases Flint, a free coding model for laptops

2Development has been largely automated through the Eureka Machine

3Flint shows good performance on benchmarks and a compression rate of 71%

4An external tester reports repeated successes on coding tasks

Why it matters : Flint illustrates a step towards self-improving AI models under human supervision, with a focus on automating development and local execution.

The startup Recursive has released Flint, a free coding model designed to run on a laptop. The company claims to have largely automated its development through its "Eureka Machine," boasting high speeds and good results on benchmarks, while ultimately aiming for human-controlled self-improvement.

Stated Goal: Towards Greater Autonomy of Agents, Under Supervision

Richard Socher presents a long-term goal of complete recursive self-improvement, with humans supervising and able to interrupt the process if necessary. Recursive asserts that each new model should enhance the Eureka Machine and accelerate the construction of the next one, estimating that if its bet succeeds, models will handle longer work periods without human assistance. Meanwhile, OpenAI indicated that by mid-August, its AI agents were logging 3.1 days of work for every day of human researcher input, signaling a broader interest in these approaches. Recursive also explained in June that its $410 million contract with AWS was less focused on headcount and more on the number of agents. On the technical side, the company mentions low-bit compression that reduced Flint by 71% with almost no loss of quality and custom kernels. In terms of usage, external tester Will Knight reports that Flint met all his coding requests during his trials.

Local Execution, Open License, and Performance Promises

Flint runs on the user's machine, allowing for offline use, private code retention, and no variable costs related to processed text. The model is offered for free on Hugging Face under the Apache 2.0 license. Technically, Flint adopts a mixture-of-experts architecture that activates only 2.5 billion of its 40 billion parameters per request, making it more compact than Meta's Muse Spark Open, announced the previous week. Muse Spark Open, in contrast, totals 816 billion parameters and requires at least one terabyte of memory. In terms of speed, Recursive reports around 131 tokens per second on a MacBook Pro M4 Pro, compared to 58 to 69 tokens per second for models like Gemma 4 and Nemotron 3.5 Lightning on the same machine, and over 150 tokens per second on Nvidia's RTX Spark. On SWE-Bench Verified, Recursive indicates that Flint outperformed other models with fewer than 50 billion parameters and that, on an average of six tests, only Alibaba's Qwen3.8 27B performed better. However, the company warns that Flint requires more guidance than large paid assistants and suggests breaking down tasks.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

What an External Tester Reports on a 48GB Mac

Will Knight reports having used Flint on a Mac equipped with 48GB of memory. He created several small programs, including a task list application and a web-scraping tool for research. He observes that smaller coding models often encounter errors or are slow compared to Claude or Codex. However, in his trials, Flint successfully responded to all his coding requests.

How Flint Was Designed with the Eureka Machine

Recursive claims that its Eureka Machine contributed to nearly every stage of Flint's development and that a small team built the model from scratch in less than four months. The tool generates ideas, conducts experiments, and decides on next steps by leveraging results, and is said to have influenced pre-training, post-training, and inference. However, the company did not specify how much time this approach saved. Flint represents an early public test of recursive self-improvement, an ambition that also raises concerns among some researchers about a loss of control. The current version described by Recursive is that of a research assistant supervised by humans, producing a model scaled for a laptop.

Who is Behind the Project and What is the Financial Support

Recursive announced in May a funding round of $650 million at a valuation of $4.65 billion, with investors such as GV, Nvidia, and AMD Ventures. The company is led by Richard Socher, former chief scientist at Salesforce. Its first model, Flint, was launched on Wednesday. It is a free coding model designed for laptop use. According to the company, Flint was largely built with the help of AI.