Recursive debuts Flint, a laptop coding model partly built by AI

Recursive has released Flint, a free coding model that runs on a laptop and was developed in large part with the help of another AI system. The launch is a notable early test of whether artificial intelligence can materially speed up the creation of better AI while still staying under human control.

The startup, which surfaced from stealth in May with a $650 million funding round and a $4.65 billion valuation, says Flint was built in less than four months by a small team using its automated research system, the Eureka Machine. The model is designed for local use on personal hardware, giving developers offline access, lower operating costs and greater privacy than cloud-based coding assistants.

Recursive’s debut arrives as the AI industry increasingly asks whether models can help design their own successors. The company’s answer is still modest compared with the most ambitious visions of recursive self-improvement, but Flint offers a public, concrete example of AI-assisted model development that investors, researchers and developers will all be watching closely.

What Recursive launched and why it matters

Recursive unveiled Flint on Wednesday as its first publicly available model, positioning it as a lightweight coding assistant that can run on a laptop rather than requiring a remote data center. That alone makes it unusual in a market where many advanced coding tools depend on proprietary cloud infrastructure and ongoing usage fees.

The launch matters for two reasons. First, it gives developers a free open model that can be used locally, privately and without per-token charges. Second, it is meant to demonstrate the company’s broader thesis: that AI research itself can be accelerated by AI systems.

Recursive says Flint is not just another small coding model. It is also a proof point for its research workflow, in which the Eureka Machine suggests ideas, runs experiments and uses the results to determine the next set of tests. In the company’s telling, the AI assistant played a role in nearly every stage of the model’s creation, from early training decisions to post-training optimization and inference tuning.

How was Flint built so quickly?

Recursive says Flint was assembled in under four months because most of the research process was automated or accelerated by the Eureka Machine. The startup describes the system as an automated research engine that can generate ideas, experiment with them and then steer subsequent work based on results.

According to the company, Eureka contributed to almost every major step in Flint’s development. That included choosing a sparser model architecture, reducing the model’s size through low-bit training and writing specialized kernels that speed up the mathematical operations underneath the model.

The company did not disclose a precise time savings figure, but the implication is clear: fewer human research hours were needed to arrive at a capable coding model. Recursive is betting that this kind of workflow will become increasingly powerful over time as each new model improves the research system that helped create it.

Recursive has said the Eureka Machine contributed across the full model-building pipeline, from pre-training to post-training and inference.

That claim places Recursive in a fast-moving corner of the AI sector where companies are trying to automate more of the research loop rather than simply automate end-user tasks.

What makes Flint different from other coding models?

Flint is designed to be compact, fast and usable on consumer hardware. Recursive says the model uses a mixture-of-experts design, which means only a subset of its full parameter set is active for any given prompt. In Flint’s case, just 2.5 billion of its 40 billion parameters are used at a time.

That matters because it allows the model to remain efficient enough for laptops while still drawing on a much larger internal capacity when needed. The result, Recursive says, is a model that is much smaller and lighter than some other recent open models intended for high-end machines.

Flint is also open source in practical terms: it is available on Hugging Face under the Apache 2.0 license, which allows free use and modification. For developers, that combination of local execution and permissive licensing is attractive because it reduces costs and makes it easier to inspect or adapt the model for specific workflows.

Why local execution is a major selling point

Local execution is a major selling point because it keeps code on the user’s machine. Developers working on proprietary products, internal tools or sensitive data may prefer that over sending prompts to a third-party cloud service.

It also changes the economics. Cloud-based coding products typically charge by usage, meaning the cost rises as developers ask more of the model. A laptop model can be used offline and without those recurring fees, which may make it especially appealing to small teams and independent engineers.

Recursive is targeting a market segment that values speed, privacy and flexibility more than raw scale. In that sense, Flint is not meant to replace the biggest coding systems on the market; it is meant to offer a more portable alternative.

How fast is Flint in practice?

Recursive says Flint is faster than several other similar-sized models on common developer hardware. On a MacBook Pro powered by Apple’s M4 Pro chip, the company says Flint produced about 131 tokens per second. That was roughly double the throughput of competing models tested on the same machine, including Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning, which Recursive said ranged from 58 to 69 tokens per second.

On Nvidia’s RTX Spark, Flint exceeded 150 tokens per second, according to the company. Those figures matter because speed is not just a technical bragging right; it affects the day-to-day usability of a model for coding workflows, where responsiveness can determine whether a tool feels fluid or frustrating.

The following table summarizes the main specs and performance claims Recursive highlighted at launch.

How does Flint compare with competing models?

Flint is not the biggest or most general-purpose model in the field, but it appears competitive in the narrow category of lightweight coding systems. Recursive says that on SWE-Bench Verified, a benchmark used to measure how well models repair real-world software bugs, Flint outperformed other models below 50 billion parameters.

In the company’s own average across six tests, Flint was beaten by only one model: Alibaba’s Qwen3.8 27B. That suggests the model is punching above its weight class, at least on the benchmarks Recursive chose to emphasize.

Flint’s scale also stands in contrast to other recent open models. Recursive pointed to Meta’s Muse Spark Open, a model released last week for Macs that reportedly uses 816 billion parameters and requires at least a terabyte of memory. By comparison, Flint is built to be much more approachable for ordinary laptop users.

This difference in size and hardware requirements highlights two distinct philosophies in AI development. One aims for maximum scale and capability, often at the cost of heavier infrastructure. The other aims for practicality, portability and lower operational cost.

Why benchmark results should be read carefully

Benchmark results should be read carefully because they do not capture every real-world coding scenario. A model may perform exceptionally on a curated test while still requiring more human guidance than larger, more expensive assistants.

Recursive itself has acknowledged that Flint needs more direction than premium paid products. The company recommends that users break bigger jobs into smaller steps, which is a reminder that local models often trade breadth for convenience and autonomy.

Still, the initial reaction appears encouraging. Early hands-on testing reported by Wired suggested the model handled small coding tasks reliably, even when smaller models often struggle with errors or lag compared with leading cloud tools.

What did early testers say?

Early testers described Flint as surprisingly capable for a laptop-friendly model. In one reported evaluation, Wired’s Will Knight used the model on a Mac with 48GB of memory to create several small applications, including a to-do list app and a web research tool.

His broader comparison was that many smaller coding models tend either to produce mistakes or to operate much more slowly than larger assistants such as Claude or Codex. By contrast, he found Flint consistently successful on the tasks he assigned.

That kind of anecdotal praise does not replace large-scale user testing, but it does matter because it suggests the model may be usable outside a lab setting. For coding tools, the real test is whether developers can trust them to complete routine tasks without constant intervention.

Early hands-on testing suggested Flint was able to complete small coding jobs reliably, even where other compact models often falter.

If those impressions hold up more broadly, Flint could become appealing to teams seeking a local assistant that is fast enough to feel modern but simple enough to run without specialized cloud access.

Why is Recursive so focused on AI improving AI?

Recursive is focused on AI improving AI because that is the core of its long-term research thesis. The company is betting that the best way to accelerate model development is to automate more of the discovery process itself, rather than relying only on human researchers to iterate manually.

The idea is often described as recursive self-improvement: an AI system helps create a better AI system, which in turn helps build an even better one. In theory, the process compounds over time.

In practice, that idea remains highly controversial. Some researchers worry that a system capable of improving itself could become difficult to predict or control. Recursive has been careful to frame its approach as human-supervised, with people able to intervene if something goes wrong.

For now, Flint is a limited example of that broader ambition. It is not a self-directing intelligence and not a system that independently redesigns itself in open-ended fashion. It is, instead, a case study in using AI to make parts of the research process faster, cheaper and more efficient.

How does this fit into the broader AI race?

This launch fits into a wider industry trend in which AI companies are using their own tools to accelerate internal research and engineering. Recursive is not alone in that approach. OpenAI recently said that by mid-August, its AI agents were contributing 3.1 workdays for every workday logged by a human researcher.

That statistic underscores a shift in the way leading AI labs think about productivity. Rather than simply asking whether AI can assist end users, companies are increasingly measuring whether AI can assist the scientists and engineers building the next generation of AI.

Recursive’s approach is also notable because of the scale of the resources behind it. The company previously disclosed a $410 million compute agreement with Amazon Web Services, which it said was less about adding human staff and more about expanding what it called “agent count.” That framing suggests the company views software agents as part of its core workforce.

In this sense, Flint is not just a product launch. It is a public demonstration of a research strategy that seeks to industrialize AI development itself.

What happens next for Recursive?

Recursive says the next stage is to make the Eureka Machine better with every model it helps create. The company’s long-term goal is to move toward full recursive self-improvement, while keeping humans in the loop to monitor and stop the process if necessary.

That ambition is still far from realization, but Flint gives the company a tangible starting point. If its AI research system can repeatedly shorten development cycles and improve output quality, Recursive could build a pipeline in which each new release speeds up the next one.

For developers, the immediate takeaway is simpler: a free, laptop-compatible coding model is now available, and it appears competitive enough to merit attention. For the broader AI industry, the more important question is whether Flint is the first of many models partly built by automated research systems.

That question will likely define how seriously Recursive’s strategy is taken in the months ahead. If the company can show that AI-assisted research produces better models faster, it could strengthen the case for automated AI development. If not, Flint may still stand as a useful open coding tool, even if the recursive dream remains ahead of schedule.

Key details at a glance

For now, Flint stands as both a product and an experiment: a local coding model designed to be useful on its own merits, and a test case for whether AI can meaningfully help build the next generation of AI systems.