Hardware Ready Ready to Bench?
Recursive has unveiled Flint, a model designed for programming that can run on a standard laptop and is available for free. The company credits its Eureka Machine with playing a significant role in the system’s development, while emphasizing that the process was conducted under human supervision.
Flint was built from scratch in less than four months thanks to the use of Eureka Machine, the automated research system developed by Recursive. The model is available on Hugging Face under an Apache 2.0 license, which permits its use and modification.
Flint by Recursive: why it matters
The project aims to bring into the real world the idea of an artificial intelligence capable of contributing to the creation of other AI systems. Recursive describes Eureka as a human-guided automated research assistant that proposes experiments and uses the results to define subsequent stages of the process.
Local execution offers practical advantages for developers: Flint can function offline, keeps programs on the user’s computer, and avoids costs associated with processing large amounts of text on cloud services. However, it requires compatible hardware resources and, according to Recursive itself, needs more detailed instructions than paid AI assistants.
Flint uses a mixture-of-experts architecture that activates only a portion of its 40 billion parameters (2.5 billion specifically) for each request. This configuration reduces computational load and makes execution possible on a common laptop, at least according to Recursive’s claims.
The comparison with Meta’s Muse Spark Open highlights differences in scale: the open-source model intended for Macs counts 816 billion parameters and requires at least one terabyte of memory. In tests conducted by the company, Flint generated approximately 131 tokens per second on a MacBook Pro with an Apple M4 Pro chip, surpassing the 58-69 tokens per second achieved by similar models like Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning on the same machine. On Nvidia RTX Spark, it would have exceeded 150 tokens per second.
On the SWE-Bench Verified benchmark, which evaluates the ability to fix real software issues, Recursive states that Flint achieved results superior to all other models with fewer than 50 billion parameters. In the average calculated over six trials, only Alibaba’s Qwen3.8 27B recorded a higher score. Complete results and execution conditions were not detailed; therefore, comparisons should be considered as preliminary indications.
What changes and what are the effects
Recursive attributes to Eureka contributions ranging from pre-training to optimization and inference, citing a sparser structure, low-precision training (which reduced Flint’s size by 71% with minimal quality loss), and custom kernels to accelerate calculations.
The startup, led by Richard Socher, former chief scientist at Salesforce, exited its private phase in May with $650 million raised and a valuation of $4.65 billion. In June, it also signed an agreement with Amazon Web Services for a commitment of $410 million in computing capacity.
Recursive considers Flint a first step toward systems capable of conducting longer research sessions with fewer human interventions, hoping that each model improves Eureka and accelerates the development of the next one, while maintaining human supervision and the possibility to interrupt the process. At present, Flint mainly demonstrates that the company has succeeded in bringing to market a local model of relatively contained size; achieving true recursive self-improvement remains a future goal.
Source and further reading on Flint by Recursive: original article.
* Content created with the assistance of artificial intelligence systems.