Recursive Flint: the open source AI coding model that runs on your Mac, even offline

Flint is an open source AI model from Recursive built for coding, and it runs on your Mac, even offline. It has 40 billion parameters, but it only uses 2.5 billion for each token. The 4-bit version takes about 20 GB of RAM: a Mac with at least 32 GB is enough. Your code stays on your computer and you don’t pay for any API.

What is Flint and who released it?

Flint is a coding language model released by Recursive on October 21, 2026. The weights are public on Hugging Face under the Apache 2.0 license, so you can download and use Flint for free, at work too.

Recursive presents it as the best in its class: small, coding-focused models that run on a regular computer. According to the official announcement, Flint leads Mixture of Experts models under 50 billion parameters on three agentic coding benchmarks: Terminal-Bench 2.1, SWE-Bench Pro v2 and SWE-Bench Verified.

How well does Flint do on benchmarks?

These are Recursive’s own figures, not my tests:

• 54.9% on Terminal-Bench 2.1, against 44.9 for Qwen 3.6 and 38.7 for Qwen 3.5 in the official chart;
• 74.2% on SWE-Bench Verified and 89.7% on SWE-Bench Pro v2, according to the model card;
• coding performance on par with dense models of 25 to 40 billion parameters, while running more than 5 times faster.

These benchmarks come from the team that built the model. Before picking it for a real project, test it on your own code.

Why is Flint so fast?

Flint is fast because it uses a Mixture of Experts architecture. Inside the model there are many "experts", meaning specialized sub-networks. For each token only a few of them switch on, the ones that are needed.

So Flint has 40 billion parameters, but only 2.5 billion do the work for each token. Less computation per token means more speed: on a MacBook Pro with an M4 Pro chip, Flint writes more than 125 tokens per second, according to Recursive.

How much RAM do you need to run Flint?

Speed has one limit: memory. Even if Flint only uses 2.5 billion at a time, all 40 billion parameters have to be loaded into RAM. At full precision, with 16 bits per parameter, that’s about 80 GB.

The fix is 4-bit quantization. Each parameter takes a quarter of the space, and the whole of Flint fits in about 20 GB of RAM. That’s why a Mac with at least 32 GB of RAM is enough to run it.

The quality loss is small. According to Recursive, their 4-bit quantized version stays within 0.7 points of the 16-bit model on the benchmarks.

How do you install Flint on a Mac?

You install Flint with LM Studio or Ollama, two of the most popular apps for running AI models locally. The steps are:

1. open LM Studio or Ollama;
2. search for "Flint";
3. download the 4-bit version;
4. connect Flint to your code editor.

If your computer doesn’t have enough RAM, you can try Flint for free in your browser, in a Hugging Face Space.

How should you use Flint for coding?

My advice is to give Flint small tasks and precise instructions. Compared with giant cloud models, a small model like Flint needs more guidance. Asking it to "rebuild the whole app" doesn’t work the way it does with bigger models. Asking it to "add validation to this form" does.

One point in Flint’s favor is context: 256,000 tokens. That’s how much text Flint reads at once, enough to hold several project files together with your instructions.

If you want to know which local model to choose for the computer you have, I compared open weight coding models across three hardware tiers. And if you wonder how well a local model holds up against cloud ones, I covered it in local AI for coding: when it’s worth it.

Who should try Flint?

Flint is worth it for anyone writing code that must not leave their computer. I’m thinking of people working on confidential company code who can’t use external APIs because of a contract or an internal policy. With Flint the model runs locally, offline, and no line of code goes through a third-party server.

If you work in one of these situations, Flint lets you use AI for coding without asking anyone for an exception. Download it, try it on a small task and see for yourself what it can do with your code.