AI startup Recursive, founded last year, now appears to be living up to its name. On Wednesday, the company released its first model, Flint. Flint is a free coding model that can run on a laptop and is reportedly built primarily with AI.
Recursive stated that its automated research system, the Eureka Machine, was involved in nearly every step of Flint's development, enabling a small team to build the model from scratch in under four months.
Flint is an early public test of "recursive self-improvement": the idea that AI can build better AI, which in turn can build even better AI. It is also one of the most controversial concepts in the industry. Some researchers worry that AI continuously improving itself could slip out of human control. For now, Recursive’s version is more limited: it is a human-supervised AI research assistant, producing models small enough to run on a laptop.
The company emerged from stealth in May of this year, raising $650 million in funding at a $4.65 billion valuation, with investors including GV, NVIDIA, and AMD Ventures. Recursive is led by CEO Richard Socher, formerly Chief Scientist at Salesforce.
Since Flint runs on the user’s own computer rather than a company’s cloud server, developers can use it offline, keep their code private, and avoid the usage fees associated with AI cloud services that charge based on the volume of text processed. The model is available for free download on the popular AI model-sharing platform Hugging Face, under the Apache 2.0 license—a permissive open-source license that allows free use and modification.
Flint is a mixture-of-experts model, meaning that when performing any given task, it only activates a small portion of itself: out of 4 billion parameters, only 2.5 billion are used at a time. This makes it lightweight enough to run on a standard laptop and significantly smaller than Meta’s open-source model Muse Spark Open, released last week for Mac. Muse Spark Open uses 816 billion parameters and requires at least 1 TB of memory.
Flint is also very fast. According to Recursive’s tests, on a MacBook Pro with the Apple M4 Pro chip, it generates approximately 131 tokens per second—about twice as fast as similar models on the same machine, including Google’s Gemma 4 and NVIDIA’s Nemotron 3.5 Lightning, which run at 58 to 69 tokens per second. On NVIDIA RTX Spark, Flint exceeds 150 tokens per second.
On the SWE-Bench Verified benchmark, which measures AI's ability to fix real software vulnerabilities, Flint outperforms all other models with fewer than 50 billion parameters. According to Recursive's average score across six tests, only Alibaba's Qwen3.8 27B scored higher than Flint.
The company acknowledges that Flint requires more guidance than large paid coding assistants and recommends breaking down large tasks into smaller steps. Nonetheless, early evaluations have been positive. Wired journalist Will Knight used Flint on a Mac with 48GB of RAM to build several small applications, including a to-do list app and a tool for web research. He found that smaller coding models often make mistakes or are slower than Claude or Codex, but Flint consistently delivered accurate results for his coding requests.
Eureka Machine generates ideas, runs experiments, and uses the results to determine what to try next. According to the company’s release video, it was involved in nearly every aspect of building Flint, from pre-training to post-training and inference.
Recursive stated that Eureka's contributions include a sparser model architecture, low-bit training—which reduced Flint's size by 71% with almost no loss in quality—and custom kernels, which are specialized programs designed to accelerate underlying computations.
Recursive did not specify how much time Eureka Machine saved researchers, but it is not the only company relying on AI for research. Last month, OpenAI stated that, as of mid-August, its AI agents had logged 3.1 times more workdays than human researchers.
If Recursive's betting pays off, each new model should be able to operate for longer periods without human assistance. The company previously told TechCrunch that its $410 million compute agreement with Amazon Web Services in June was “less about adding employees and more about adding agents.”
Socher told Wired that, in the long term, the goal is for Eureka Machine to achieve full recursive self-improvement, with humans overseeing the entire process and able to halt it if issues arise. Recursive stated that each new model should make Eureka Machine a better researcher, enabling it to build the next model faster.