Free coding model for the laptop was largely developed by AI itself

The young AI company Recursive released its first AI model on 21 October. The model is called Flint and helps people write computer code. It is free and runs on an ordinary laptop, so without an internet connection and without your code going to another company. What makes it special is how it was made: according to Recursive, its own AI system did a large part of the development work.

Recursive was founded last year and is led by Richard Socher, who was previously head of research at Salesforce. The company wants to build AI that keeps making itself better. In May it came out with 650 million dollars (about 580 million euros) in investment, from GV, Nvidia and AMD Ventures among others. That valued the company at 4.65 billion dollars (about 4.1 billion euros).

An AI that helps build AI

Recursive's system is called the Eureka Machine. It does what a researcher does: it comes up with ideas, carries them out, tests whether they work and, based on the outcome, chooses what to try next. It explores several directions at once and combines ideas that look promising. People supervise it.

According to Recursive, the Eureka Machine helped with almost every step: choosing the training data, the way of training and making the model run fast. This allowed a small team of people and AI agents to build Flint from scratch in less than four months. A few examples the company gives:

• Building the model. The system tried out different designs and chose the design that promised the best combination of speed and programming quality.
• Making it smaller. After training, the system stored the numbers in the model in a shorter, coarser form, roughly like saving a photo at lower quality. According to Recursive, this made Flint 71 per cent smaller, while the test scores stayed almost the same.
• Faster calculation. The system wrote special calculation programs tailored to Apple and Nvidia chips, so that the model runs faster on them.

Large model, little calculation

Flint is a so-called mixture-of-experts model. It contains 40 billion parameters, the adjustable numbers in which an AI model stores its knowledge. For each step it uses about 2.5 billion of them.

Because Flint needs little calculation per step, it fits on a laptop. For comparison: the open coding model Muse Spark Open from Meta, which appeared last week, has 816 billion parameters and needs at least a terabyte (1,000 gigabytes) of working memory.

Flint is also fast. On a MacBook Pro with Apple's M4 Pro chip, it produces more than 125 tokens (small pieces of words) per second according to Recursive, faster than you can read along. That is about twice as fast as comparable models on the same laptop, such as Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning. On an Nvidia RTX Spark, a small AI computer for the desk, Flint reaches more than 150 tokens per second.

How well does it program?

Recursive ran Flint on six tests. A well-known one is SWE-Bench Verified: a collection of real bugs from existing software, which an AI has to fix on its own. According to Recursive, Flint scores better on it, and on two other programming tests, than all other models with the same design and fewer than 50 billion parameters. Over the average of the six tests, only Alibaba's Qwen3.8 27B scores higher. That is a model that uses all its parameters at once. According to Recursive, Flint programs about as well as such models of 25 to 40 billion parameters, but is more than five times as fast.

These figures come from Recursive itself. Because anyone can download the model, researchers can now check them themselves.

A first impression came from journalist Will Knight of Wired. Using Flint on a Mac with 48 gigabytes of memory, he built a few small programs, including a to-do list and a helper that searches the internet for research. Smaller coding models often got stuck or were slow for him, but according to Knight, Flint carried out every task well.

Getting started yourself

Flint is available for free on Hugging Face, a popular website for sharing AI models. It falls under the Apache 2.0 licence: anyone may use it, adapt it and build it into their own applications. It works with common programs for running AI models locally, such as Ollama and LM Studio. Recursive is also releasing its own program that makes Flint run extra fast on the newest Macs.

Recursive is open about what Flint needs. The model requires more guidance than the large paid coding assistants. The company recommends splitting big jobs into smaller steps and giving clear instructions. Anyone who does that has, according to Recursive, a fast coding helper that works entirely on their own computer.

What does this mean for the road to the singularity?

Flint is an early, public test of an idea that many AI researchers are watching: AI that builds better AI, which in turn builds better AI again. This is called recursive self-improvement. If such a cycle gets going, progress can go faster and faster. Some researchers are worried that such a process could get out of human control. Recursive therefore builds step by step, with people watching who can step in.

The current version is still modest: a research AI under human supervision that delivered a small, fast model. But it shows that AI is already involved in almost every part of building an AI model, from training to making it calculate faster on a laptop. Recursive's goal is that every new model makes the Eureka Machine a better researcher, so that the next generation is ready sooner. In June, the company signed a deal worth 410 million dollars (about 365 million euros) with Amazon Web Services for computing power. The company told TechCrunch at the time that this is more about the number of AI agents than the number of employees.

Recursive is not alone in this. OpenAI reported in September that by mid-August its AI agents were already doing 3.1 working days of research work for every working day of a human researcher. In this way, AI is becoming a full colleague in the AI lab. Recursive's coming models will show how fast that cycle can start turning, and Flint already gives programmers a free foretaste on their own laptop.

What is the technological singularity? Read the explanation →