Flint model tests self-improving AI with open programming on laptops

The self-improving AI test starts from a laptop screen, away from closed cloud computing servers. Recursive launched its first programming-focused model, Flint, an open-source, free model developed in under four months by a small human team that relied on an almost fully automated research system the company calls the Eureka Machine. This offering provides a practical test of what the tech community describes as iterative self-improvement, where algorithms build systems that surpass their own capability, which then create something better than themselves.

The new model relies on a mixture-of-experts architecture, activating only 2.5 billion parameters at any given moment out of a total of 40 billion parameters in the system. This lightweight structure makes the model runnable locally on an ordinary personal computer, putting it in stark contrast with giant open models such as Meta’s recent Muse Spark Open, which comprises 816 billion parameters and requires a full terabyte of memory to operate. According to the company’s tests, Flint generated 131 tokens per second on a MacBook Pro equipped with a M4 Pro chip, roughly twice the speed of competing models on the same device, such as Google’s Gemini 4 and Nvidia’s Neuron 3.5 Lightning, and its speed exceeded 150 tokens per second on a desktop RTX Spark card.

The driving force behind the project lies in the algorithms’ ability to engineer more efficient programming tools with minimal human intervention.Recursive, led by former Salesforce chief scientist Richard Sutherland, which went public in May with $650 million in funding and a valuation of about $4.65 billion from investors including GV, Nvidia and AMD’s venture arm, describes the achievement as a direct result of its automated research system. The Eureka Machine is involved in every stage of Flint’s construction, from initial training through subsequent refinement and inference processing; the engine delivered a more modular architecture, employed low-bit training that reduced the model’s size by 71 percent without any noticeable loss in performance accuracy, and added custom software kernels to accelerate computation, concretely reflecting the company’s strategy after it signed a $410 million cloud-computing deal with Amazon Web Services designed to increase the number of intelligent agents rather than human staff.

In efficiency benchmarks, Flint outperformed the SWE Bench Verified metric designed to test models’ ability to handle real programming errors compared with models under 50 billion parameters, and the only model that surpassed it in the average of six company-run tests was Alibaba’s Qwen 3.8, which has 27 billion parameters. Although the field trial documented by Wired on a computer with 48 GB of memory demonstrated its success in building small, fast experimental programs without the typical slowdown of miniature models, the company acknowledged that the model requires more human guidance than paid cloud-based coding assistants, advising developers to break large tasks into smaller programming steps to ensure execution accuracy.

This technological shift recalibrates the software infrastructure calculations for development teams in the Gulf, Egypt and the Levant.Making a model this fast available under the Apache 2.0 license and distributing it via Hugging Face resolves two long-standing challenges for regional companies: cumulative inference costs and the risk of exposing sensitive source code. Instead of sending banking or government programming data through foreign cloud interfaces that are subject to complex security and audit requirements, your team can run the coding assistant entirely offline within the local work environment, without a single line of code leaving the device. In budget terms, this solution shifts outsourcing firms and emerging software companies from paying escalating token-consumption bills to relying on developers’ existing computers, while requiring you as a developer to reshape your skills toward architectural task decomposition rather than depending on generic commands aimed at closed cloud assistants.

The launch of Flint is a tangible indicator of a reshuffling of scientific research roles, a trend also noted by major labs such as OpenAI, which recently announced that its agents log 3.1 research-day equivalents for every one research-day performed by a human researcher. While the ambition of achieving full iterative self-improvement still raises concerns about loss of control and demands strict human oversight, delivering a local tool capable of handling real development tasks demonstrates that shrinking models and tuning their efficiency through AI itself constitutes a practical path to competing with giant cloud monopolies.