Recursive Introduces Flint: An AI-Driven Laptop Coding Model
Recursive, an AI startup launched last year, is starting to live up to its name. The company, which aims to develop self-improving AI, introduced its first model, Flint, on Wednesday. Flint is a free coding model that can run on a laptop and was largely created with the help of AI.
The creation of Flint benefited from Recursive’s automated research system, known as the Eureka Machine, which played a crucial role in nearly every stage of development. This allowed a small team to build the model from scratch in under four months.
Flint acts as a preliminary public experiment in recursive self-improvement: the theory that AI can produce improved AI, leading to even more sophisticated AI. This idea is one of the most debated in the sector, as some researchers fear that self-enhancing AI may evade human control. Currently, Recursive’s method is more structured: an AI research assistant overseen by humans has generated a model compact enough to operate on a laptop.
Emerging from stealth mode in May, Recursive obtained $650 million in funding, achieving a valuation of $4.65 billion from backers like GV, Nvidia, and AMD Ventures. The company is led by CEO Richard Socher, who previously served as the chief scientist at Salesforce.
Since Flint runs on the user’s local machine rather than on external cloud servers, developers can utilize it offline, safeguard the confidentiality of their code, and avoid the expenses linked to cloud-based AI services that charge by text processing volume. The model is available for free download on Hugging Face, a widely-used platform for sharing AI models, under the Apache 2.0 license, allowing for unrestricted use and modification.
Flint is a mixture-of-experts model, meaning it utilizes only a fraction of its full capacity for certain tasks: applying just 2.5 billion of its 40 billion parameters at any given moment. This efficiency makes it lightweight enough for typical laptops, significantly smaller than Meta’s Muse Spark Open, an open model for Macs that was released last week with 816 billion parameters and requiring at least a terabyte of memory.
Flint also showcases remarkable speed. In Recursive’s tests, it generated about 131 tokens (text segments) per second on a MacBook Pro equipped with Apple’s M4 Pro chip, nearly doubling the performance of similar models, including Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning, which achieved 58 to 69 tokens per second on the same machine. On Nvidia’s RTX Spark, Flint surpassed 150 tokens per second.
In benchmarks such as SWE-Bench Verified, which evaluate how effectively AI can resolve real-world software bugs, Flint outshined other models with fewer than 50 billion parameters. In Recursive’s own assessments across six tests, only one model, Alibaba’s Qwen3.8 27B, scored higher than Flint.
The company acknowledged that Flint would need more guidance compared to larger, paid coding assistants and suggested breaking down complex tasks into smaller parts. However, initial feedback has been positive. Wired’s Will Knight used Flint on a Mac with 48GB of memory to build various small applications, including a to-do list app and a tool for conducting online research. He noted that smaller coding models often struggled with errors or slow performance compared to Claude or Codex, whereas Flint handled his coding requests successfully each time.
The Eureka Machine generates ideas, performs experiments, and assesses the results to decide on the next steps. According to the company’s launch video, it was instrumental in every phase of Flint’s development, from pretraining to post-training and inference.
Recursive highlighted that Eureka’s contributions included a more effective model structure, low-bit training that reduced Flint’s size by 71% with minimal drop in quality, and custom kernels, specialized programs that improve computational performance.
While Recursive did not specify the time saved by the Eureka Machine, it reflects a broader trend of increasing AI usage in research. OpenAI reported last month that its AI agents logged 3.1 workdays for each human research day by mid-August.
If Recursive’s approach is successful, each new model should be able to handle longer tasks with reduced human input. The company previously informed TechCrunch that its $410 million computing agreement with Amazon Web Services in June was “less about headcount and more about agent count.”
Looking ahead, Socher told Wired that the ultimate goal is for the Eureka Machine to achieve full recursive self-improvement, with human supervision in place to halt the process when needed. Recursive claims that each subsequent model should enhance the capabilities of the Eureka Machine, accelerating the development of the next generation.
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.


