Can You Run a Free AI Coding Model on Your Laptop? Recursive Flint

Hey hey hey, welcome to Agentic Pod! I'm Liam.

And I'm Stella! Liam, a free coding model just dropped that writes 131 tokens a second on a MacBook Pro. Is that real, or is that a marketing number?

It's the company's own number, so we'll poke at it. It's called Flint, it came out October 21st, and it runs on your laptop. No API bill, no cloud.

Free, offline, and fast. Okay, I have questions, and one of them is "what's the catch."

Start simple. Who made it?

A startup called Recursive. The CEO is Richard Socher, he used to be chief scientist at Salesforce. And here's the fun part, a small team built it in under four months, with an AI research system of their own doing a lot of the work.

So an AI helped build the coding AI.

Yep, they call that system the Eureka Machine. Humans still supervised it and picked which results to keep.

Okay, and the size? Forty billion parameters sounds like it needs a data center.

Here's the trick. It's a mixture of experts model. About forty billion parameters in total, but only around two and a half billion wake up for each token.

So it's big on paper, small in practice.

Right, like a huge office where only the three people you need pick up the phone. That's why it's fast. And it's Apache 2.0, so you can download it, change it, use it at work.

Fast how? Compared to what?

Recursive says about 131 tokens a second on a MacBook Pro with an M4 Pro chip. Over 150 on an Nvidia RTX Spark. Similar models on that same MacBook ran at roughly 58 to 69.

So double. Again, their numbers.

Their numbers, yes. Now, running it. The easy way is Ollama. One line in your terminal: ollama run flint colon 40b.

And how big is that download?

Twenty-three gigabytes. So no, a sixteen gig laptop is not going to be happy. The Wired reviewer who tried it used a Mac with 48 gigabytes of memory.

Sad for my little laptop. Is there another way?

Yeah. There are builds on Hugging Face from RecursiveSI, an INT4 one for Apple Silicon and an NVFP4 one for other GPUs. The context window on the Ollama page says 262 thousand tokens.

Okay, but I don't want to chat with it. I want it to do real work.

Then use it the way Recursive tells you to. Small steps. They admit Flint needs more explicit instructions than the big paid assistants.

So no "build me a startup" prompt.

Ha, no. Try this. Step one: "Write only the tests for a function that slugifies titles. Stop after the tests." Step two: "Now write the function so those tests pass." Step three: "Explain what could break."

Three small asks instead of one giant one.

Exactly. Each step is easy to check, and a local model never gets a chance to wander off. And it's private. Your code never leaves your laptop, which matters for client work.

And it's free, so I can make it redo a step ten times and nobody sends me a bill.

Right. Ollama also serves it on your own machine, so editors and tools that can talk to a local Ollama model can use it. Check your tool's docs for that.

Okay, so what's the honest catch?

Three. One, the benchmarks are Recursive's own. They say it beat other models under fifty billion parameters on SWE-bench Verified, and across six tests only Alibaba's Qwen3.8 27B scored higher.

So second place, by their own scorecard.

Which is still impressive for a free model. Two, speed is not quality. Token speed tells you how fast it types, not how well it codes. Three, that Wired reviewer only tried small projects, a to-do app and a web research tool. Big codebases are unproven.

So who's it for?

Developers with a 48 gig Mac or a strong GPU, people who care about privacy or hate usage bills, and anyone who wants to experiment. If you need the very best on a giant codebase, keep your paid tool open next to it.

My verdict. Install it on a weekend, give it small jobs, and watch it.

Same. Links to the sources and the guide are in the description. Which job would you hand to a free local model first? Tell us in the comments. See you next time!