AI startup Recursive has introduced Flint, an open-weight coding model designed to run on local computers, in a development that offers an early look at how artificial intelligence can help build other AI systems. The company says its automated research platform, called Eureka Machine, contributed to nearly every stage of Flint’s development, from model architecture and training methods to inference optimization. A small team developed the model from scratch in less than four months, according to Recursive.
Flint is built for coding agents that can inspect software projects, identify bugs, implement features and write tests. Its mixture-of-experts architecture contains approximately 40 billion parameters but activates only 2.5 billion for each token, helping reduce computational demands during inference. The model is available for download under the Apache 2.0 license, enabling developers to run it on compatible hardware without relying exclusively on a cloud-based AI service. For Recursive, Flint is also a test of a broader ambition: using AI to automate more of the research process that produces future AI models.
Key Takeaways
- New coding model: Recursive has released Flint, an open-weight model designed for agentic software development.
- Built with AI assistance: The company says its Eureka Machine contributed throughout model development.
- Developed in under four months: A small team of people and AI agents built the model from scratch.
- 40 billion parameters: Flint activates approximately 2.5 billion parameters per token through its mixture-of-experts architecture.
- Local execution: Developers can run the model on compatible personal computers and supported GPU hardware.
- Fast inference: Recursive reports speeds above 125 tokens per second on an M4 Pro MacBook Pro.
- Open license: Flint’s weights are available under Apache 2.0.
- Long-term ambition: Recursive wants to build AI systems that can progressively improve the process of developing AI.
What Is Recursive Flint?
Flint is a coding-focused language model developed by Recursive, an AI startup working on automated AI research and model improvement.
The model is intended for agentic coding, a form of software development in which an AI system does more than suggest lines of code. A coding agent can inspect an existing repository, plan changes, use tools, run tests and attempt to correct errors.
These capabilities can help developers automate parts of the software development process. For example, an agent might investigate why a test is failing, trace the problem to a particular function, propose a fix and run tests to check whether the change works.
Recursive says Flint performs strongly among mixture-of-experts models with fewer than 50 billion total parameters on several coding benchmarks. Its published evaluations include SWE-Bench Verified, SWE-Bench Pro v2 and Terminal-Bench 2.1.
These are company-reported benchmark results rather than independent guarantees of performance on every software project. Real-world results depend on the complexity of the task, the surrounding tools, the model configuration and the quality of the instructions it receives.
How Flint Was Built With AI
The most distinctive part of Flint’s launch is not just the coding model itself, but how Recursive says it was developed.
The company used an automated research system called Eureka Machine to explore possible improvements, run experiments and evaluate results. Rather than asking researchers to test every idea manually, the system can help generate candidate approaches, implement experiments and use their outcomes to decide what to investigate next.
What Eureka Machine contributed
According to Recursive, the system contributed to several important parts of Flint’s development:
- Model architecture: It explored different configurations to balance coding capability and inference speed.
- Training methods: It helped investigate training recipes and optimization techniques.
- Quantization: It searched for ways to reduce the model’s memory footprint while preserving performance.
- Inference optimization: It developed specialized computational kernels designed to improve execution on supported hardware.
- Evaluation: It helped compare experimental configurations and identify promising approaches.
This approach can shorten the feedback loop between an idea and a tested result. Researchers can use AI agents to explore more alternatives, while people continue to supervise the process and assess whether the resulting improvements are useful.
Recursive says the work allowed its small team to build Flint in under four months. The company has not publicly quantified precisely how much development time was saved compared with a fully human-led process.
Flint’s Mixture-of-Experts Architecture Explained
Flint contains approximately 40 billion parameters, but it does not activate all of them for every token it generates. Instead, it uses a mixture-of-experts, or MoE, architecture that routes work through selected parts of the model.
This design can reduce the computation needed for each generated token compared with a dense model of similar total size, although actual performance depends on the architecture, memory bandwidth and inference software.
Flint technical specifications
| Feature | Details |
|---|---|
| Developer | Recursive |
| Model name | Flint |
| Primary use | Agentic coding |
| Architecture | Mixture of experts |
| Total parameters | Approximately 40 billion |
| Active parameters per token | Approximately 2.5 billion |
| Context window | Up to 512,000 tokens supported by the model configuration |
| License | Apache 2.0 |
| Availability | Downloadable model weights |
| Local hardware | Compatible Apple Silicon and supported GPU systems |
| Reported M4 Pro speed | More than 125 tokens per second |
A long context window can help a coding model work with larger repositories and longer task histories. However, a model’s maximum context capacity is not a guarantee of uniform accuracy across every position or task. Developers still need to manage context, tools and task complexity carefully.
Flint Targets Developers Who Want Local AI Coding
Many coding assistants rely on remote servers, where model inference is billed through subscriptions or usage-based APIs. Flint offers another option: running a coding model on hardware controlled by the developer.
Local execution can provide several benefits. Developers may be able to work offline, reduce dependence on external services and keep source code on their own machines. The open license also allows organizations to adapt the model for their own applications, subject to the license terms.
However, local deployment does not mean every computer can run the model comfortably. The full model weights require substantial memory unless a quantized version is used, and performance varies with hardware and inference software.
Recursive reports that Flint can generate more than 125 tokens per second on a MacBook Pro with an M4 Pro chip under its stated test conditions. It also reports speeds above 150 tokens per second on supported NVIDIA RTX Spark hardware.
Those figures are useful indicators of potential responsiveness, but they are not universal results. Prompt length, context size, model precision, software configuration and the type of task can all affect speed.
The company also acknowledges that Flint may require more explicit guidance than larger, closed-source coding assistants. Breaking complex assignments into smaller steps can help developers obtain more reliable results.
Why Recursive Wants AI to Build Better AI
Recursive’s broader objective is to automate more of AI research and development. This idea is often described as recursive self-improvement: an AI system helps develop a more capable successor, which can then contribute to further improvements.
Flint represents an early practical test of that approach. Eureka Machine helped develop the model, and Recursive hopes that more capable future models will make its research system more effective.
The potential benefit is a faster development cycle. If AI systems can independently propose useful experiments, implement them and assess the outcomes, researchers may be able to explore more design choices with the same resources.
That does not mean Flint independently created itself or that Recursive has achieved fully autonomous AI research. People designed the overall system, established the objectives and supervised the work. The company’s longer-term vision involves automating more of those steps while retaining human oversight.
There are also unresolved questions about how reliably such systems can evaluate their own results. Automated experiments can produce misleading conclusions if the evaluation process is flawed, and apparent benchmark improvements may not translate into better performance outside controlled tests.
Funding and the Competitive Landscape
Recursive emerged from stealth in May 2026 with $650 million in funding at a reported valuation of $4.65 billion. Its backers include GV, Nvidia and AMD Ventures, according to reporting on the company.
The funding gives Recursive resources to pursue a demanding research agenda involving model development, computing infrastructure and automated experimentation.
Flint also enters a competitive market for local and open-weight coding models. Developers can choose among models from multiple AI companies, with differences in coding quality, speed, memory requirements, licensing and hardware compatibility.
Recursive is positioning Flint around a particular combination: coding performance, local inference and development accelerated by its automated research system. Its ability to sustain that advantage will depend on independent evaluations, adoption by developers and the quality of future releases.
The Bigger Picture
Flint is notable for combining an open-weight coding model with a development process that relies heavily on AI-assisted research. If systems such as Eureka Machine can reliably automate model experimentation, they could help smaller teams investigate more ideas and improve model performance without scaling their research staff at the same rate. Local execution also offers developers more control over their code and infrastructure.
However, the broader promise of AI building AI remains a work in progress. Flint’s capabilities and speed claims need to be evaluated against real-world coding tasks, while recursive self-improvement raises questions about evaluation quality, oversight and how much of the research process can be automated reliably. The launch is an early demonstration of the approach, not proof that AI research can already operate independently from human direction.
FAQs
1. What is Recursive Flint?
Flint is an open-weight AI coding model designed to help coding agents inspect codebases, fix bugs, implement features and write tests. It is designed to run locally on compatible hardware.
2. What does it mean that Flint was largely built by AI?
Recursive says its Eureka Machine contributed to nearly every stage of development, including architecture exploration, training, quantization and inference optimization. Human researchers remained involved in designing and supervising the process.
3. Can Flint run on a laptop?
Yes, on compatible hardware. Recursive reports that Flint can run on an M4 Pro MacBook Pro, but memory requirements and performance depend on the model version, quantization and software setup. Not every laptop will be suitable.
4. Is Flint free to use?
The model weights are available under the Apache 2.0 license, which generally permits use, modification and redistribution subject to its terms. Running the model still requires compatible hardware and computing resources.
5. What is recursive self-improvement in AI?
It is the idea that AI systems can help research, develop and improve subsequent AI systems. Recursive is pursuing this approach through Eureka Machine, with Flint serving as an early example of AI-assisted model development.
Looking Ahead
Recursive’s next steps will likely focus on improving its automated research system, releasing more capable models and demonstrating that AI-assisted development can consistently produce better results. Flint gives the company a public product through which developers can evaluate its approach to local coding, while the model’s open license allows broader experimentation and adaptation.
The larger question is whether each generation of AI research tools can make the next generation meaningfully easier to develop. If Recursive can demonstrate repeatable improvements in coding performance, training efficiency and research productivity, its approach could influence how future models are built. For now, Flint is a promising early test of that strategy, with independent evaluations and continued releases likely to determine how much practical value it delivers.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
