A year-old AI startup called Recursive has released its first publicly available model, a coding assistant named Flint. The company, which is focused on developing AI systems capable of improving themselves, launched Flint on Wednesday. The model is free, designed to run locally on a laptop, and was built with significant involvement from Recursive's own automated research platform.

Recursive describes its internal research system as the Eureka Machine. According to the company, this system contributed to nearly every stage of Flint's development, allowing a small team to take the model from concept to completion in fewer than four months.

A Practical Test of Self-Improving AI

Flint represents an early, real-world demonstration of recursive self-improvement - the concept that AI systems can be used to build better AI, which can then be used to build even better AI after that. It is also one of the most contested ideas in the field. A number of researchers have raised concerns that AI systems capable of improving themselves could become difficult or impossible for humans to manage. For now, Recursive's approach remains tightly scoped: the Eureka Machine functions as an AI research assistant operating under human supervision, and the model it helped produce is compact enough to run on consumer hardware.

Recursive came out of stealth in May, announcing $650 million in funding at a valuation of $4.65 billion. Investors include GV, Nvidia, and AMD Ventures. The company is led by CEO Richard Socher, who previously served as chief scientist at Salesforce.

Why Local Matters

Because Flint runs directly on a user's machine rather than through a company's cloud infrastructure, it offers several practical advantages. Developers can work offline, keep their source code entirely private, and avoid the token-based usage fees that cloud-hosted AI services typically charge. The model is available for free download on Hugging Face, the widely used AI model-sharing platform, under the Apache 2.0 open-source license, which permits free use, modification, and redistribution.

Flint is built as a mixture-of-experts model, an architecture that activates only a portion of its total parameters for any given task. Specifically, it draws on just 2.5 billion of its 40 billion parameters at a time. This design makes the model lightweight enough for a standard laptop, and stands in sharp contrast to larger open models. For comparison, Meta's recently released Muse Spark Open - an open model built with Mac hardware in mind - uses 816 billion parameters and requires at least one terabyte of memory to run.

Performance Benchmarks

In internal testing, Flint demonstrated notable speed. On a MacBook Pro equipped with Apple's M4 Pro chip, the model generated approximately 131 tokens per second. That figure is roughly double the output of comparable models tested on the same hardware, including Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning, which produced between 58 and 69 tokens per second. On Nvidia's RTX Spark hardware, Flint exceeded 150 tokens per second.

On SWE-Bench Verified, an established benchmark that evaluates how effectively AI models can identify and fix real software bugs, Flint outperformed all other models with fewer than 50 billion parameters. Across Recursive's own aggregate scoring of six separate tests, only one competing model scored higher than Flint: Alibaba's Qwen3.8 27B.

Recursive did acknowledge that Flint may require more user guidance than the larger, subscription-based coding assistants on the market. The company recommends breaking complex tasks into smaller, more manageable steps to get the best results. Early external testing has been generally favorable. A review by technology journalist Will Knight, writing for Wired, found that Flint handled every coding task he gave it correctly. Knight ran the model on a Mac with 48GB of memory and used it to build several small programs, including a to-do list application and a web research tool. He noted that smaller coding models often struggle with errors or sluggish performance compared to tools like Claude or Codex, but found that Flint did not exhibit those problems.

How the Eureka Machine Contributed

The Eureka Machine works by generating research ideas, running experiments, and using the results of those experiments to determine the next course of action. According to Recursive's launch materials, it played a role in nearly every phase of Flint's construction, spanning pre-training, post-training, and inference optimization.

Among the specific contributions Recursive attributes to the Eureka Machine are a more efficient model architecture, a low-bit training technique that reduced Flint's size by 71 percent with minimal impact on performance quality, and the development of custom computational kernels - specialized software routines that accelerate the underlying mathematical operations the model relies on.

Recursive has not disclosed how much time the Eureka Machine saved compared to a fully human-led research process. The company is not alone in using AI to accelerate AI research, however. OpenAI stated last month that its AI agents were generating the equivalent of 3.1 working days of research output for every single day worked by a human researcher, a figure it expected to reach by mid-August.

The Road Ahead

Recursive's broader ambition is for each successive model to take on progressively longer and more complex research tasks with less human involvement. The company's $410 million compute agreement with Amazon Web Services, announced in June, was described by Recursive as being driven less by the need for human employees and more by the need for AI agents to run at scale.

Socher told Wired that the long-term objective for the Eureka Machine is to reach full recursive self-improvement, with humans remaining in a supervisory role and retaining the ability to intervene or halt the process if something goes wrong. The company's position is that each new model it builds will make the Eureka Machine a more capable researcher, enabling it to build the next model more efficiently than the one before it.