Founded last year, AI startup Recursive is starting to live up to its name. The company, which aims to build AI that can improve itself, released its first model, Flint, on Wednesday. Flint is a free coding model that runs on your laptop and was built primarily with the help of AI.
Recursive’s automated research system is called Eureka Machine, and the company says it was instrumental in nearly every stage of Flint’s development. This allowed a small team to build a model from scratch within four months.
Flint is an early public test of recursive self-improvement. So the idea is that AI will build better AI, and that AI will build even better AI. It’s also one of the most discussed ideas in the industry. Some researchers worry that AI could improve itself and move out of human control. For now, Recursive’s version is more limited, an AI research assistant overseen by people who have created models small enough to fit on laptops.
The company came out of stealth in May when it raised $650 million in funding at a $4.65 billion valuation from backers including GV, Nvidia, and AMD Ventures. Recursive is led by CEO Richard Socher, former chief scientist at Salesforce.
Because Flint runs on users’ own computers rather than on corporate cloud servers, developers can use it offline, keep their code private, and avoid usage fees that AI cloud services charge based on the amount of text their models process. The model is available for free download on Hugging Face, a popular site for sharing AI models, under the Apache 2.0 License, an open source license that allows free use and modification.
Flint is a mixed model of experts. That is, it uses only a small slice of itself for a specific task. Of the 40 billion settings called parameters, you only use 2.5 billion at a time. As such, it’s light enough to be a regular laptop, and much smaller than Meta’s Muse Spark Open, an open model built for Mac that was released last week. Muse Spark Open uses 816 billion parameters and requires at least 1 terabyte of memory.
Flint is also fast. In Recursive’s tests, a MacBook Pro with Apple’s M4 Pro chip generated approximately 131 tokens (text chunks) per second. This is roughly double the 58-69 tokens per second of similar models managed on the same machine, such as Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning. On Nvidia’s RTX Spark, Flint topped 150 tokens per second.
In SWE-Bench Verified, a benchmark that measures how well AI can fix bugs in real-world software, Flint outperformed other models under 50 billion parameters. In Recursive’s own average across six tests, only one model, Alibaba’s Qwen3.8 27B, scored higher than Flint.
The company acknowledged that Flint needed more guidance than the big paid coding assistants, and suggested breaking down large tasks into smaller steps. Still, early reviews were positive. Wired’s Will Knight used Flint on a Mac with 48 GB of memory to build several small programs, including a to-do list app and a tool to search and explore the web. Although he found that smaller coding models often had errors and were slower than Claude or Codex, Flint fulfilled his coding requests every time.
The Eureka Machine comes up with ideas, runs experiments, and uses the results to decide what to try next. According to the company’s announcement video, he contributed to nearly every part of building Flint, from pre-training to post-training to inference.
Recursive said Eureka’s contributions include sparser model shapes, low-bit training that reduced Flint by 71% with little loss in quality, and custom kernels, specialized programs that speed up the underlying computations.
Recursive won’t say how much time Eureka Machine saves researchers, but it’s not the only company relying on AI for research. OpenAI announced last month that its AI agents were logging 3.1 work days for every human researcher worked by mid-August.
If Recursive’s bet pays off, each new model should be able to withstand long hours of work without human assistance. The company previously told TechCrunch that its $410 million computing contract with Amazon Web Services in June was “more about agent numbers than headcount.”
The long-term goal is for the Eureka machine to achieve full recursive self-improvement, with humans able to oversee the process and stop it if something goes wrong, Sorcher told Wired. Recursive says that with each new model, the Eureka machine will become a better researcher and be able to build the next model faster.
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
