Flint is a free coding model from Recursive that runs on laptops, using just 2.5 billion parameters out of 40 billion for each task. Its key advantages are offline use, keeping code on the user’s device, and no usage-based text charges.
The company says Eureka Machine, an automated research system, helped develop Flint through nearly every stage, including model training and speed optimization. The important thing to know is that Flint is still better suited to breaking large tasks into smaller steps and requires more guidance than paid coding assistants.
Flint on a Laptop
Flint is a coding model designed to run on laptops, making it suitable for developers who want a coding assistant installed locally without constantly relying on the company’s servers.
The model uses the Apache 2.0 license, which allows free use and modification. Anyone interested can download Flint from Hugging Face and test it on their own machine.
Imagine writing code on a computer without an internet connection, or working on a project whose code you do not want to send elsewhere. Flint runs directly on the user’s computer, helping protect privacy and avoid usage-based charges.
But the key question is how well a model small enough for a laptop can actually write code, because convenience means little if results are slow or errors appear throughout the process.
Where Flint Fits into Recursive’s Plans
Flint is the first model from Recursive, an AI company founded in 2025. It is also an early public test of the idea of using AI to help create AI, with humans providing oversight.
The company raised $650 million at a valuation of $4.65 billion, with Richard Socher, former chief scientist at Salesforce, serving as CEO.
No Previous Version to Compare, but Flint’s Starting Point Is Clear
Available sources do not disclose any previous version of Flint, so we should not create a comparison table between an old and new version ourselves. What can be identified is the Eureka Machine system and the development process the company used before launching this first model.
| Factor | What existed before launch | Flint |
|---|---|---|
| What is mentioned | Eureka Machine and the development process | Free coding model |
| Use | Generating ideas and conducting experiments | Runs on laptops |
| Status | System used to create the model | Available for download |
What It Is Like to Use in Different Situations
If you need to write code on your own machine and do not want to upload it to the cloud, Flint can run offline. This makes it suitable for work that requires keeping code private, without paying based on the number of messages.
People who want to download or modify the model can do so because Flint uses the Apache 2.0 license. For interactive coding, Recursive says it can reach approximately 131 tokens per second on a MacBook Pro with an M4 Pro chip.
For debugging, Flint performed better than other models with fewer than 50 billion parameters on SWE-Bench Verified. In Wired’s experiments, it also successfully built a to-do app with web research tools every time.
Flint Compared with the Models Already Available to Users
Flint stands out for laptop use, while Muse Spark Open requires at least 1 terabyte of memory. Gemma 4 and Nemotron 3.5 Lightning achieve speeds of 58–69 tokens per second on a MacBook Pro with an M4 Pro chip.
| Factor | Flint | Alternative models |
|---|---|---|
| Muse Spark Open | Works on laptops | 816 billion parameters and requires at least 1 terabyte of memory |
| Speed on a MacBook Pro with an M4 Pro chip | Approximately 131 tokens per second | Gemma 4 and Nemotron 3.5 Lightning: 58–69 tokens per second |
| Average score across 6 tests | Higher than Qwen3.8 27B | Qwen3.8 27B scored higher than Flint |
Clear Strengths and Acceptable Limitations
Flint is suitable for people who want to write code locally without connecting to the cloud. It is free, runs offline, and helps keep code private. Its limitations are that it still requires more detailed instructions than paid tools, and large tasks should be divided into smaller steps.
Pros
- +Free to download and use
- +Runs on laptops and works offline
- +Keeps code on the device, reducing privacy concerns
- +No usage fees based on the amount of text processed
- +Responsive for coding tasks
Cons
- −Requires more guidance than paid coding tools
- −Large tasks should be divided into smaller steps
- −Still an early test of this concept
Costs That Are Not on the Price Tag
Flint is free to download and does not require cloud usage fees based on the amount of text processed. However, you still need a machine with enough memory to run the model locally. One review used a Mac with 48GB of memory.
Another cost is the time required to use it. Large tasks should be divided into smaller steps, and Flint requires more guidance than paid coding tools. You may therefore need to spend more time checking code, refining instructions, and testing results yourself. It is best suited to people willing to trade time for privacy and lower long-term costs.
Who Flint Is For, and Who Should Wait
Made for
- Developers who want to write code offline
- People who prioritize privacy and open models
- Laptop users who work in smaller steps
Think twice
- People working on large, continuous tasks who do not want to keep dividing them
Skip this one
- People who want a paid assistant that requires less guidance — consider paid coding tools
The More Important Test Than Speed Is Relying on AI to Create AI
Flint is therefore important not only because it runs on laptops, but also because it is an early test of how much AI can help design, train, and develop the next generation of models. Recursive says that Eureka Machine contributed to nearly every stage of Flint’s creation.
The next question is how long Eureka Machine can continue operating and how much it can reduce reliance on human researchers while still keeping people in control and able to stop the system when problems occur. This is a direction worth following just as closely as Flint’s benchmark scores.