Machine-Made Code: Recursive Debuts Flint, a Lighter Laptop Coding Model Built by AI

Artificial intelligence research startup Recursive released its first open-source model, Flint, showcasing a practical milestone for self-improving software pipelines. Led by Chief Executive Officer Richard Socher, former chief scientist at Salesforce, the firm developed Flint using an automated internal research engine dubbed the Eureka Machine. The automated setup directed key stages across the model’s design, training, and optimization, allowing a small human engineering team to build a custom 40-billion-parameter coding model from scratch in less than four months.

Flint operates as a mixture-of-experts model, activating just 2.5 billion parameters per token during code generation. Activating a small subset of parameters keeps memory requirements low, allowing software developers to run the entire system directly on standard laptops, MacBooks, or personal workstations without sending proprietary code to central cloud servers. Running model inference locally preserves source code privacy, eliminates recurring API usage charges, and provides offline functionality for remote software developers.

During internal speed benchmarks, Flint processed roughly 131 tokens per second on an Apple MacBook Pro equipped with an M4 Pro chip. That generation rate doubles the output speed of competing open-weight models tested on identical laptop hardware, including Meta’s Muse Spark Open, Google’s Gemma 4, and Nvidia’s Nemotron 3.5 Lightning. On dedicated desktop hardware like Nvidia’s RTX Spark, Flint reached processing speeds topping 150 tokens per second.

On software benchmarks, Flint outperformed several competing models under 50 billion parameters. On SWE-bench Verified, an industry standard evaluating real-world bug fixes, Alibaba’s Qwen model was the only sub-50B competitor to score higher than Flint in average testing. Early technical reviews confirmed strong performance across everyday development tasks, including generating small utility applications, setting up web crawlers, and fixing code syntax errors, though reviewers noted that smaller models still require clear step-by-step instructions when handling large, complex codebases.

The Eureka Machine automated critical stages throughout Flint’s four-month build cycle. The automated research system proposed model architectures, ran experiments, evaluated training recipes, and wrote custom low-bit execution kernels tailored specifically for Apple Silicon and Nvidia chips. These automated contributions reduced low-bit training steps by 71 percent without degrading code quality or logic reasoning capabilities.

Recursive emerged from stealth earlier this year after securing $650 million in funding at a $4.65 billion valuation, backed by prominent investors including Google Ventures, Nvidia, and AMD Ventures. The company previously booked $410 million in cloud compute access with Amazon Web Services, opting to invest heavily in automated research agents rather than expanding human developer headcount. Other AI developers are adopting similar strategies; for example, OpenAI recently revealed that its internal research agents log 31 workdays of automated experiments for every single workday logged by a human engineer.

Socher emphasized that while the Eureka Machine automates daily experimental workflows, human researchers remain in full control of project goals, safety guardrails, and final validation. Releasing Flint under the open-source Apache 2.0 license gives the public free access to download, test, modify, and deploy the weights for commercial or personal development work. Demonstrating that automated research systems can design fast, efficient local models establishes a practical framework for expanding automated software development across the broader tech industry.