Recursive released Flint, an open-weight coding model, on October 21, 2026. Much of the research behind the model was carried out with the help of AI, and the entire model was built with the assistance of a small team in less than four months. The open question is whether this approach creates a chance for small laboratories to compete with the industry's biggest players without huge investment. For readers following that question, the release is a test case: it pairs an unusually small team with an unusually automated research process, and both parts of that equation can be checked over time.

Flint uses a mixture-of-experts design, activating only a small part of the model at a time, which helps it run efficiently. Its model card lists 39.8 billion parameters, with about 2.5 billion active per token. The Apache 2.0 license allows commercial use, and developers can run Flint locally. That openness is the practical hook: anyone can download the weights, inspect them, and run the model on their own machine without an API key or a metered bill. The sparse mixture-of-experts design is also what makes local use realistic, since only about 2.5 billion parameters do the work for any given token.

What Recursive says the Eureka Machine built

Recursive credits its Eureka Machine with suggesting, coding, and testing changes to Flint's architecture, training, and inference software. CEO Richard Socher told WIRED that AI improved "nearly every aspect of the entire model pipeline."

That approach is central to Recursive's stated goal of having AI systems help build better successors. Socher maintains that people can still stop the research process. If the claim holds up, the implication for the wider industry is about where model-development effort goes: less of it in routine research engineering, and more of it in the compute-heavy work of running experiments.

Company benchmarks and a WIRED hands-on

In tests conducted by Recursive, Flint outperformed other mixture-of-experts models with fewer than 50 billion parameters on SWE-Bench Pro v2, Terminal-Bench 2.1, and SWE-Bench Verified. The company also reported speeds above 125 tokens per second on an M4 Pro MacBook Pro and 150 tokens per second on an NVIDIA RTX Spark at short context lengths.

One important consideration: these are company numbers, and independent evaluation is still limited. WIRED's Will Knight tested Flint himself and found it quite adept, even able to create a small website. However, a handful of successful coding tasks do not show how it compares with other models.

OpenHands conducted tests on Flint before its launch and provided developers access to it on the day of rollout. While that indicates Flint is capable of real-world coding tasks, OpenHands has not independently assessed Recursive's benchmarking claims. That gap is the clearest thing to watch next: the benchmarks are public, so third-party results would show whether the company's numbers hold outside its own test setup.

The $410 million question about cheap AI

Creating a model that requires fewer people is no guarantee that it will be cheap. Recursive has not revealed how much Flint cost to build. On July 28, 2026, the company entered a $410 million deal with AWS, as reported by TechCrunch. It previously raised $650 million, and none of these amounts show the cost of building Flint itself. For the small-lab question, this is the complication to keep in view: automation can reduce headcount, but the spending behind the model still runs into the hundreds of millions.

Research from Epoch AI indicates why expenses stay high. Final training runs made up 9.6% of OpenAI's research-compute expenditures in 2024, with the majority of funding distributed to other research-related activities.

AI agents can take over some jobs previously handled by researchers. However, running experiments and testing new ideas still requires a large amount of computing power. That is why stories about tiny teams building models and stories about enormous compute budgets are not contradictions: the labor input and the compute input tell two different parts of the cost story.

Open weights, local hardware, and global competition

Meanwhile, using AI is becoming much cheaper. AI estimates that the cost of achieving comparable performance has fallen about 47% per quarter since 2023.

On September 22, 2026, OpenAI introduced GPT-6 Sol and Luna at prices at least 50% below their predecessors, as Cryptopolitan reported. Flint had nothing to do with those price changes. Still, it gives developers another option: a coding model they can run on their own hardware without relying on paid cloud services.

The benefits could reach beyond established AI markets. Stanford's 2026 AI Index found that open-source development is helping more countries participate in AI innovation. A permissively licensed, locally runnable coding model fits that pattern, since it puts capability in the hands of developers who cannot or prefer not to pay for frontier cloud access.

For now, Flint shows what a small team can accomplish with AI-assisted research. Whether the same approach can make model development affordable for smaller labs remains an open question. The near-term signals to watch are independent benchmark results, disclosure of what the build actually cost, and how much of Recursive's next model the Eureka Machine ends up writing.