Recursive Flint AI Coding Model: Laptop Speed and Claims

What Recursive Says the Eureka Machine Did

Recursive says its automated research system, called the Eureka Machine, contributed across nearly every stage of Flint’s development. According to the company, the system helped with work involved in researching, building and improving the model. That makes Flint a test of Recursive’s broader approach: using AI systems to assist in creating other AI systems.

The company’s claims do not mean Flint was developed without people. Wired’s coverage describes human oversight as part of Recursive’s approach to recursive self-improvement. The idea is to use AI to accelerate research and development while people remain involved in guiding and reviewing the work.

That distinction matters. “Built largely by AI” describes the company’s account of how it developed Flint; it does not, by itself, establish how much work was automated or how the system’s contributions compare with those of its human researchers.

Flint’s Mixture-of-Experts Design and Reported Laptop Speed

TechCrunch reports that Flint uses a mixture-of-experts architecture with 40 billion parameters in total, while activating 2.5 billion for a given task. In this type of design, different parts of a model can be selected for different inputs, rather than using every parameter for every response. The stated aim is to balance capability with lower computing demands.

Recursive reports that Flint reached about 131 tokens per second on a MacBook Pro with an M4 Pro chip. That figure comes from the company’s own testing, so it should be treated as a reported result rather than an independently verified benchmark. Real-world speed can vary with the computer, software setup, task and measurement method.

The laptop focus sets Flint apart from much larger models. TechCrunch contrasts it with Meta’s Muse Spark Open, described as an 816-billion-parameter open model requiring at least one terabyte of memory. The comparison illustrates different hardware targets, but the figures alone do not show which model performs better on coding tasks. Model size, memory requirements and token speed are only part of the picture.

What Flint’s Launch Claims Do - and Don’t - Establish

Wired reports that it tested Flint, adding an outside perspective beyond Recursive’s own product claims. However, the available reporting does not provide a complete independent benchmark comparison establishing how Flint performs against other coding models across a broad set of tasks. The reported laptop speed also should not be confused with measures of code quality or reliability.

For developers, Flint’s appeal will depend on more than whether it runs locally. Useful questions include how well it handles different programming languages, whether its suggestions work in real projects, what hardware and setup are needed, and how its results compare with other tools. The launch establishes Recursive’s product and its design goals; broader testing will help clarify how well those goals translate into everyday use.

Recursive is positioning Flint as both a local coding model and evidence of its effort to use AI in AI development. Follow our coverage of AI coding models and the systems being developed to help build them.

Frequently Asked Questions