Recursive releases Flint, a laptop-ready coding model built by AI

Recursive, an AI startup led by former Salesforce chief scientist Richard Socher, has released its first public model, Flint. Launched on Wednesday, this free coding assistant is designed to run entirely on local hardware, such as a standard laptop, rather than relying on cloud infrastructure. The release marks a significant step in the company’s mission to develop AI systems capable of improving themselves through automated research.

What happened

Flint is an open-weight coding model available for download on Hugging Face under the Apache 2.0 license. This licensing choice allows developers to use, modify, and distribute the model freely. By running locally, the model enables users to keep their code private, work offline, and avoid the token-based usage fees common with cloud-hosted AI services. Recursive emerged from stealth earlier this year with substantial backing, including $650 million in funding at a $4.65 billion valuation from investors like GV, Nvidia, and AMD Ventures.

The development process itself was heavily automated. Recursive reports that its internal automated research system, called the Eureka Machine, assisted in nearly every stage of Flint’s creation. This assistance allowed a small human team to build the model from scratch in less than four months. The company positions Flint as an early practical test of recursive self-improvement, a concept where AI systems help design and train subsequent, more capable AI models. While some industry observers worry about the safety implications of self-improving AI, Recursive maintains that humans currently oversee the Eureka Machine and can intervene if necessary.

How it works

Flint utilizes a mixture-of-experts architecture, a design that activates only a subset of the model’s parameters for any given task. Although Flint contains 40 billion parameters in total, it uses only 2.5 billion at a time. This sparsity makes the model significantly lighter and faster than dense models of similar total size. For comparison, Meta’s recently released Muse Spark Open model uses 816 billion parameters and requires at least a terabyte of memory, whereas Flint is optimized for ordinary consumer laptops.

The Eureka Machine contributed specific technical optimizations to achieve this efficiency. According to Recursive, the system helped design a sparser model shape and implemented low-bit training techniques that reduced the model’s size by 71% with minimal loss in quality. It also developed custom kernels, which are specialized software routines that accelerate the underlying mathematical computations. These innovations allow Flint to generate text rapidly on consumer hardware without sacrificing too much performance.

Key details

• Flint generates approximately 131 tokens per second on a MacBook Pro with an Apple M4 Pro chip, roughly double the speed of comparable models like Google’s Gemma 4.
• On Nvidia’s RTX Spark hardware, the model achieved speeds exceeding 150 tokens per second.
• In SWE-Bench Verified tests, which measure bug-fixing capabilities, Flint outperformed other models with fewer than 50 billion parameters.
• The model was built in under four months with significant assistance from Recursive’s automated Eureka Machine system.
• Flint is free to download and use under the Apache 2.0 open source license.
• Early independent testing by Wired’s Will Knight showed the model successfully completed coding tasks on a Mac with 48GB of memory.

Why it matters

For software engineers and technical founders, the shift toward high-performance local models reduces dependency on third-party API providers. Running code generation locally eliminates latency issues associated with network requests and ensures that proprietary codebases never leave the developer’s machine. This is particularly valuable for teams working in regulated industries or those handling sensitive intellectual property who cannot use public cloud AI services.

Furthermore, Flint demonstrates the accelerating impact of AI-assisted research. If automated systems like the Eureka Machine can indeed compress development cycles from years to months, the pace of innovation in open-source AI may increase dramatically. However, developers should note that local models still require more explicit guidance than large cloud-based assistants. Recursive advises breaking complex tasks into smaller steps to get the best results, indicating that while the technology is powerful, it is not yet a fully autonomous replacement for senior engineering judgment.

What you can do

• Download Flint from Hugging Face to test its performance on your local hardware setup.
• Refactor large coding prompts into smaller, discrete tasks to improve the model’s accuracy and reliability.
• Compare Flint’s token generation speed against your current local models using standard benchmarking tools.
• Review the Apache 2.0 license terms to ensure compliance with your organization’s open source usage policies.
• Experiment with offline workflows to identify which parts of your development pipeline can operate without internet access.
• Monitor updates from Recursive regarding the Eureka Machine to understand how future iterations might further automate model optimization.