Recursive Launches Flint, a Free Laptop Coding Model

Flint brings free, private coding AI to laptops Recursive’s first model, Flint, is a free coding assistant designed to run on a user’s laptop rather than a company’s cloud servers. That means developers can use it offline, keep their code private and avoid the usage fees charged by cloud AI services based on how much text a model processes. Flint is available to download on Hugging Face under the Apache 2.0 license, an open source license that allows free use and modification. Its laptop-friendly design relies on a mixture-of-experts architecture: for any given task, Flint activates just 2.5 billion of its 40 billion parameters. Using only a portion of the model at a time helps make it light enough to run on an ordinary laptop. That local approach offers an alternative to sending code to remote services, while the free license lets users modify the model. Flint is an early public test of Recursive’s effort to build AI that can help develop better AI. The company says its automated research system, the Eureka Machine, helped with nearly every step of Flint’s development, and that a small team built the model from scratch in under four months. Flint’s speed and coding benchmark results Recursive says Flint is fast on both Apple and Nvidia hardware. In the company’s tests, the model generated about 131 tokens per second on a MacBook Pro with an M4 Pro chip—roughly twice the 58 to 69 tokens per second achieved by similar models, including Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning, on the same machine. On Nvidia’s RTX Spark, Flint exceeded 150 tokens per second. On SWE-Bench Verified, which measures how well AI can fix real-world software bugs, Flint outperformed other models with fewer than 50 billion parameters. Across Recursive’s own average of six tests, only Alibaba’s Qwen3.8 27B scored higher. Recursive cautions that Flint may need more guidance than large, paid coding assistants, recommending that users break big tasks into smaller steps. Early feedback has been encouraging, though: Wired’s Will Knight tested Flint on a Mac with 48GB of memory, asking it to build several small programs, including a to-do list app and a web research tool. While Knight noted that smaller coding models can be slow or error-prone compared with Claude or Codex, he reported that Flint handled his coding requests every time. How Eureka Machine helped build Flint Recursive describes the Eureka Machine as an automated research system: it proposes ideas, runs experiments, then uses the results to decide what to try next. The company says it helped with nearly every step of Flint’s development, from pretraining to post-training and inference, enabling a small team to build the model from scratch in under four months. Its reported contributions include designing a sparser model shape and developing low-bit training that reduced Flint’s size by 71% with almost no loss in quality. Eureka also helped create custom kernels—specialized programs that speed up the computations beneath the model. Together, these efforts supported Recursive’s goal of making Flint small enough to run on a laptop. The system is an AI research assistant overseen by people, rather than a fully autonomous researcher. Recursive has not said how much time Eureka saved its researchers. The company’s longer-term aim is for each new model to make the Eureka Machine a better researcher, so it can help build the next model faster. CEO Richard Socher has described the eventual goal as full recursive self-improvement, with humans supervising the process and able to stop it if something goes wrong. Recursive’s plans for AI-led research Recursive emerged from stealth in May with $650 million in funding at a $4.65 billion valuation, backed by firms including GV, Nvidia and AMD Ventures. The startup is led by CEO Richard Socher, Salesforce’s former chief scientist, and is working to build AI that can improve itself. Its Eureka Machine is an automated research system that proposes ideas, runs experiments and uses the results to decide what to try next. For now, Recursive describes its approach as a human-supervised research assistant: people oversee the process and can stop it if something goes wrong. Its longer-term ambition, Socher told Wired, is full recursive self-improvement, with each new model helping Eureka become a better researcher and build the next model faster. The company said its $410 million compute deal with Amazon Web Services, signed in June, was “less about headcount and more about agent count.” The goal is for each new model to handle longer stretches of work without human help. Recursive is not alone in assigning research tasks to AI agents: OpenAI reported that by mid-August its agents were logging 3.1 workdays for every one worked by a human researcher.