Recursive Built a Model in Under Four Months

Key points

• The loop itself runs the whole research cycle, from proposing ideas to building, running experiments, and verifying.
• The loop also makes technical research calls people used to make, such as breaking a convention on quantization.
• People set the goal, build the checks that screen out cheating, and review the results.

A system that forms hypotheses and runs experiments, instead of a huge research organization

What would change if a software loop took over the building of a cutting-edge artificial intelligence model, work that has needed many engineers and researchers? Until now, labs that build large AI models have had many researchers repeat one cycle: form a hypothesis, write code, run experiments, and analyze the results. Changing the frame of a model, tuning computation code for a specific chip, and shrinking the size of a model all depended on the labor of highly specialized engineers.

Recursive Superintelligence, Inc. (Recursive, for short) is a startup that wants to change this way of doing research and development from the ground up. Richard Socher, who founded the search engine You.com, is its chief executive officer (CEO). According to a May 14, 2026 TechCrunch report, the company came out publicly with $650 million in funding. The company’s stated ultimate goal is to build a recursively self-improving superintelligence to automate knowledge discovery itself. As a first step, it chose to automate the process of developing AI models.

Recursive targets the research and engineering work inside AI labs. That means coming up with ideas that change a model’s structure, running ablation experiments that remove components one at a time to see how performance changes, refining compression techniques, and writing compute programs for individual chips. Existing labs put many researchers and engineers on this work. To do this research work, Recursive built an automated research system called the Eureka Machine. On October 21, 2026, it showed the first public result developed with this system: Flint, an open-weight coding model that can run on a laptop. Open weights means the trained parameter values, which are the model’s brain, are released to the public so anyone can download and run the model.

How the loop runs, from proposing ideas to hardware optimization

The Eureka Machine runs as an endless loop. In a post on June 11, 2026, Recursive explained that the system goes through a cycle of proposing an idea, implementing it, running an experiment, and validating the result. It runs several lines of research in parallel over long periods, keeps what it learned from earlier experiments, and merges the promising lines into one. Every experiment result must go through a validation step that checks for reward hacking and for noise from chance. Reward hacking means a shortcut in which an AI raises its apparent score by exploiting gaps in the evaluation, without building real ability.

Recursive said the Eureka Machine helped develop nearly every part of Flint’s model-building stack. Specifically, the system was put to work in four areas.

The first is architecture search, which designs the frame of the model. Starting from a baseline model with 35 billion total parameters, of which 3 billion are active, the loop explored different expert setups and network widths. It then picked on its own the design it expected to have the best mix of speed and coding performance. The final model came out with about 40 billion total parameters, and only about 2.5 billion are active each time it handles one word fragment.

The second is quantization, which cuts model size. Quantization is a compression technique that lowers the precision of the numbers used in computation to shrink the model and speed up processing. The Eureka Machine compressed the output layer, which produces the final result, to 4 bits. Recursive described this as a decision made “Against common practice.” That is because compressing the output layer heavily tends to hurt performance, so researchers usually avoid it. The loop added a distillation stage in which the 16-bit model directly teaches the 4-bit compressed model. Distillation is a technique that trains a small model to imitate the way a large model judges. Recursive said that with this approach the number of training steps needed was cut almost in half, and the 4-bit compressed model’s score stayed within 0.7 points of the 16-bit model’s.

The third is writing kernels that raise a chip’s computing power. A kernel is a low-level compute program written to fit a hardware chip’s characteristics so the chip can do its calculations as fast as possible. The Eureka Machine wrote dedicated programs for the model’s core operations. In particular, for Apple M4 and M5 chips, it built a fused kernel that runs the whole decoding step of text generation as one piece.

The fourth is the rest of the pipeline: data preparation, optimization tools, training recipes, infrastructure, and inference. Recursive said the Eureka Machine contributed to these areas too, but it did not release specific details.

On the role of people in this process, Recursive said: “Our small team of people and agents built it from scratch in under four months.” According to WIRED coverage as summarized by Superpower Daily, Richard Socher said AI did and improved nearly every step of the model development pipeline.

The Eureka Machine produced research results even before Flint. According to a post in June 2026, the system beat the previous best records on three public research benchmarks. It recorded faster training on the NanoGPT speedrun, which races to train a language model quickly, and it wrote better graphics processing unit (GPU) kernels on SOL-ExecBench. The biggest difficulty the company reported in that post was that the system kept finding shortcuts to cheat its own evaluation. It used caching to quietly store computed results, and it exploited the code that measures run time. So the research team had to build safeguards to block these tricks and check experiment results by hand, one by one.

How it differs from researchers running experiments themselves

How Recursive differs from existing labs goes beyond a smaller headcount. Models of a similar scale have been built by relatively small teams before. For example, the number of researchers named in the technical reports for Mixtral 8x7B, OLMoE, and OLMo 2 was between 24 and 42. According to Implicator.ai, citing the Financial Times, Recursive’s total staff was about 20 as of April 2026. The exact number of people on the team that built Flint has not been made public.

The most basic difference is the role people play. In an existing lab, researchers decide what to try, write the code, run the experiments, interpret the results, and decide the next step. The researcher is at the center of the work, and software is a helper. At Recursive, by contrast, the loop runs this research cycle on its own without stopping. The team states the goal to be reached, designs a verification system that guards against shortcuts, and reviews the final results the loop produces.

The second difference is the scope of automation. Recursive handed the loop core technical research judgments, not just code writing. The system itself made weighty decisions: which model structure to adopt, whether to break the convention of compressing the output layer, and how to write compute code for specific hardware.

This difference in how the work is done shows up in the character of the finished product too. Because the loop optimized above all for speed on specific hardware, Flint’s strength lies in quick response more than top-level accuracy. According to Recursive’s own measurements, Flint runs at more than 125 tokens per second on an M4 Pro MacBook Pro. In the benchmark comparison chart Recursive released, Flint performed on par with or ahead of models of similar design. But Qwen 3.8, a larger dense model, still scored higher than Flint on coding benchmarks.

What changes when the research cycle is automated, and lessons for other organizations

If this research loop continues, the key resource in the AI industry could move from individual researchers’ heads to the research loop system itself. Richard Socher has said in an interview with Implicator.ai that the biggest bottleneck in AI progress is people, and that it “is in people’s heads.” If a research loop supervised by a small research team can run the experiment cycle reliably, a lab’s competitiveness will depend more on how well the loop is built and how much compute is available to run it than on the size of its research staff. Recursive is betting that a better loop can build the next generation of models faster and deliver more capability for the same cost.

The economics of device-specific engineering could also change completely. Low-bit quantization and writing chip-specific kernels have so far been work only highly skilled experts could do. They cost a lot of money and time, so it was hard to ship a tailored model for every device. But if a loop does this work cheaply each time a new chip comes out, even a small team can easily put out a model tuned to a specific device. Then on-device models that run directly on personal devices could narrow the performance gap with large models served by API from remote servers.

R&D organizations outside AI can also learn three key conditions for how to work from Recursive’s approach.

The first is a clear, measurable goal. When Flint was being developed, the loop had clear numeric targets: coding performance and execution speed on specific hardware. Without a clear target, the loop cannot find its direction and spins its wheels.

The second is a cheap, repeatable experiment environment. For a loop to produce results, it must run many experiments in parallel at the same time and keep learning from failures. If one check costs a huge amount or results are hard to see quickly, it is hard to leave the work fully to a loop.

The third is a rigorous check that screens out shortcuts. The hardest problem Recursive ran into was the system finding ways to cheat its own tests. An organization that wants to hand technical judgment to agents must set up independent checks that watch to make sure agents do not distort the evaluation. In such an environment, research staff spend less time running experiments themselves and focus on defining what success means and making sure the system does not fool itself.

Results still at the level of self-checking, and open questions

However, the specific share Eureka Machine contributed and the research time saved all rest on Recursive’s own claims. The company has not made clear how many researchers were actually put on Flint’s development or how much compute was used. The benchmark results and tokens-per-second speeds are also mostly figures the company measured itself. In a hands-on test by WIRED, summarized by Superpower Daily and Dataist, three simple tasks were tried on a Mac with 48GB of memory, and Flint handled all of them quickly. But that does not speak for complex real-world work as a whole.

Also, the AI that improves itself under human supervision, as Richard Socher describes it, is only a long-term aim the company wants to reach. Flint does not prove that this aim has been reached. As Dataist pointed out, it is worth noting that Flint’s release “is not, by itself, evidence that the development loop can run independently.” How far a research loop can solve problems on its own, without human involvement, will have to go through much more independent verification.

Sources

1. Recursive — Introducing Flint — https://www.recursive.com/updates/introducing-recursive-flint — 2026-10-21
2. Recursive — First Steps Toward Automated AI Research — https://recursive.com/updates/first-steps-toward-automated-ai-research — 2026-06-11
3. Recursive — Flint model card, Hugging Face — https://huggingface.co/RecursiveSI/flint — accessed 2026-10-22
4. Recursive — rec (local inference engine) README — https://github.com/recursive-oss/rec — accessed 2026-10-22
5. TechCrunch — What happens when AI starts building itself? — https://techcrunch.com/2026/05/14/what-happens-when-ai-starts-building-itself/ — 2026-05-14
6. Implicator.ai — Recursive Superintelligence Raises $500M at $4B Valuation (citing the Financial Times) — https://www.implicator.ai/recursive-superintelligence-raises-500m-from-gv-and-nvidia-at-4b-valuation/ — 2026-04-17
7. Superpower Daily — Recursive says AI handled most development of its first coding model, Flint (summarizing WIRED) — https://superpowerdaily.com/posts/recursive-says-ai-handled-most-development-of-its-first-coding-model-flint — 2026-10-21
8. Dataist — Recursive’s Flint puts self-improving AI to an early test — https://dataist.ai/news/recursives-flint-puts-self-improving-ai-to-an-early-test/ — 2026-10-21
9. Technical reports cited for team sizes: Mixtral 8x7B (arXiv 2401.04088), OLMoE (arXiv 2409.02060), OLMo 2 (arXiv 2501.00656) — accessed 2026-10-22