Recursive Releases Flint, a Free Laptop Coding Model Built With AI-Assisted Research

TL;DR

• Recursive launched Flint, a free coding model developed from scratch in under four months with help from its Eureka Machine research system.
• The model has 40 billion parameters but activates only 2.5 billion per task through a mixture-of-experts architecture.
• Flint is available on Hugging Face under the Apache 2.0 licence and supports local, offline use.

Recursive Introduces Its First Public Model

AI startup Recursive released Flint on Wednesday, introducing a coding model designed to run locally on a laptop.

The company says a small team built Flint from scratch in less than four months, using its automated research system, the Eureka Machine, throughout much of development.

The release provides an early public test of Recursive’s ambition to use AI to accelerate the creation of better AI systems. For now, the process remains supervised by human researchers.

Recursive emerged from stealth in May with $650 million in funding at a $4.65 billion valuation. Its backers include GV, Nvidia and AMD Ventures. CEO Richard Socher previously served as Salesforce’s chief scientist.

Flint Offers Local Coding Without Cloud Usage Charges

Flint is free to download from Hugging Face under the Apache 2.0 licence, which permits use and modification subject to its terms.

Running the model locally allows developers to work offline and avoid the per-token charges associated with cloud AI services. It also lets them keep code on their own machines, although privacy still depends on any external tools or services connected to the model.

Flint uses a mixture-of-experts architecture with 40 billion total parameters, activating approximately 2.5 billion for a given task.

That selective activation reduces computational demands. However, the full model’s storage and memory requirements still matter; the active parameter count alone does not determine which laptops can run it effectively.

Company Tests Show Strong Speed and Coding Results

Recursive reported that Flint generated approximately 131 tokens per second on a MacBook Pro equipped with Apple’s M4 Pro chip.

In the company’s tests, comparable models—including Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning—produced between 58 and 69 tokens per second on the same machine. Flint exceeded 150 tokens per second on Nvidia’s RTX Spark, Recursive said.

The company also reported that Flint outperformed other models below 50 billion parameters on SWE-bench Verified, which evaluates the ability to resolve real-world software issues.

Across Recursive’s own average of six evaluations, only Alibaba’s Qwen3.8 27B ranked higher.

These are company-reported comparisons. Results can vary with hardware, model configuration, testing methods and the software used to run each system.

Early Testing Highlights Capabilities and Limits

Recursive acknowledged that Flint may require more guidance than larger paid coding assistants. It recommends dividing complex projects into smaller, manageable tasks.

An early hands-on report from Wired’s Will Knight was positive. Using a Mac with 48GB of memory, Knight tested Flint on several small programs, including a to-do list application and a web research tool, and reported successful results.

Those examples suggest practical potential for smaller development projects, but they do not establish consistent performance across larger codebases or more demanding engineering work.

Eureka Machine Contributed Across Development

Recursive says the Eureka Machine proposes research ideas, runs experiments and uses their results to guide subsequent work.

Its contributions to Flint reportedly spanned pretraining, post-training and inference optimization.

The company highlighted a sparser architecture, low-bit training that reduced the model’s size by 71% with little reported quality loss, and custom computational kernels designed to improve execution speed.

Recursive did not quantify how much researcher time the system saved, leaving its precise productivity contribution unclear.

Self-Improving AI Remains the Longer-Term Goal

Recursive’s broader objective is for each successive model to improve the Eureka Machine’s research abilities, helping it develop the next model more efficiently.

The company previously described its $410 million computing agreement with Amazon Web Services as supporting growth in AI agents rather than simply expanding human headcount.

Socher told Wired that the eventual aim is full recursive self-improvement, with humans overseeing the process and retaining the ability to stop it.

Flint represents a narrower milestone: an openly available coding model produced through substantial AI assistance, with human supervision still central to development.