AI Startup Recursive was founded last year, and it now seems to be living up to its name. The company released its first model, Flint, on Wednesday. Flint is a free coding model that can run on laptops and is said to be primarily built using AI.

Recursive indicates that its automated research system Eureka Machine was involved in almost every step of the development process of Flint, which enabled a small team to complete this model from scratch in less than four months.

Flint is an early public test of "recursive self-improvement": that is, AI can build a better AI, which in turn can build an even better AI. This is also one of the most controversial ideas in the industry. Some researchers are concerned that the continuous self-improvement of AI may get out of human control. For now, the version of Recursive is more limited: it is a AI research assistant supervised by humans, and the models it produces are small enough to run on laptops.

The company ended its stealth mode in May this year, raised $650 million in financing, and was valued at $4.65 billion. Investors included GV, NVIDIA, and AMD Ventures. Recursive is led by CEO Richard Socher, who previously served as the chief scientist at Salesforce.

Since Flint runs on the user's own computer, rather than on the company's cloud server, developers can use it offline, maintaining the privacy of their code and avoiding the costs associated with using the AI cloud service, which charges based on the amount of text processed by the model. This model can be downloaded for free from the popular AI model sharing website Hugging Face, and it is licensed under the Apache 2.0 license, which is an open-source license that allows for free use and modification.

Flint is a hybrid expert model, which means that when performing any given task, it only utilizes a small portion of its capabilities: out of 4 billion parameters, only 2.5 billion are used each time. As a result, it is light enough to run on an ordinary laptop and is much smaller than the open-source model Muse Spark Open that was released last week for Mac. Muse Spark Open uses 816 billion parameters and requires at least 1TB of memory.

The speed of Flint is also very fast. According to tests by Recursive, on the MacBook Pro equipped with Apple’s M4 Pro chip, it can generate approximately 131 token (text blocks) per second, which is about twice as fast as similar models on the same machine. The latter models include Google’s Gemma 4 and NVIDIA’s Nemotron 3.5, which have a speed of 58 to 69 token per second. On NVIDIA RTX Spark, the speed of Flint exceeds 150 token per second.

In the benchmark SWE - Bench Verified that measures the ability of AI to fix real software vulnerabilities, Flint outperforms other models with a parameter volume of less than 50 billion. According to the average score of Recursive in six tests, only Alibaba's Qwen3.8 with a score of 27B surpasses Flint.

The company acknowledges that Flint requires more guidance than larger, paid coding assistants and suggests breaking large tasks down into smaller steps. Nevertheless, early evaluations were positive. Wired journalist Will Knight used Flint to build several mini-programs on a Mac equipped with 48GB of memory, including a to-do list application and a tool for web research. He found that smaller coding models often made mistakes or were slower compared to Claude or Codex, but Flint was able to accurately complete his coding requests every time.

Eureka Machine comes up with ideas, runs experiments, and uses the results to decide what to try next. According to the company's release videos, it was involved in almost every aspect of building Flint, from pre-training to post-training to inference.

Recursive indicates that the contributions of Eureka include a sparser model structure, low-bit training – which reduced the size of Flint by 71% with almost no loss in quality – as well as a custom kernel, which is a specialized program used to accelerate underlying computations.

Recursive does not specify how much time Eureka Machine has saved for researchers, but it is not the only company that relies on AI for its research. OpenAI stated last month that as of mid-August, the number of working days recorded by its AI agents was 3.1 times that of human researchers.

If the bet on Recursive pays off, each new model should be able to work for longer periods of time without human assistance. The company previously told TechCrunch that the $410 million computing power agreement it reached with Amazon Web Services in June "was more about increasing the number of agents than about adding employees."

In an interview with Wired, Socher stated that in the long run, the goal is to enable Eureka Machine to achieve complete recursive self-improvement, with humans overseeing the entire process and being able to stop it when issues arise. Recursive argued that each new model should make Eureka Machine a better researcher, thereby helping it to build the next model more quickly.