Recursive Flint: what it is and how to run it locally

Recursive Flint is a free, open-weight coding model that runs on your own computer instead of in someone else's cloud. It came out on October 21 under the Apache 2.0 license, which means a business can download it, use it commercially and pay nothing per token.

Recursive CEO Richard Socher announced it on X as the company's first model, built with its automated research system, the Eureka Machine. His pitch fits in one line: "Private, unmetered intelligence on your own machine." Within the hour, @dr_cintas was calling it "huge for local AI coding" and noting it works with Ollama, LM Studio and llama.cpp.

What Flint is

Flint is a mixture-of-experts model. The model card on Hugging Face lists 39.8 billion total parameters, but only about 2.5 billion do work on any given token. That's why it can be fast on a laptop: the whole model sits in memory, but each step touches only a small slice.

The rest of the specs, from Recursive:

• Context window: 262K tokens on the model card. Recursive's launch post says it supports more but recommends staying within 256K.
• Speed: more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 on a 64GB NVIDIA RTX Spark, at short contexts.
• Benchmarks: 54.9% on Terminal-Bench 2.1 and 74.2% on SWE-bench Verified, per Recursive's own evaluations.

All of those numbers come from Recursive. @StockSavvyShay noted the 125 tokens per second figure came from "short-context tests," and Recursive's own footnotes say so too. Nobody has published independent tests yet.

Recursive admits one weakness. The launch post says Flint "may need more guidance" than the big closed coding models and suggests splitting complex jobs into smaller subtasks with explicit instructions. I'd take that at face value: plan the big jobs yourself and hand Flint the pieces.

How to run Flint

Which route fits depends on the machine on your desk.

Ollama

The Ollama v0.40.8 pre-release adds Flint. Install it, then run:

ollama run flint

The release notes say Flint is supported on macOS for now, with other GPU support coming in a later release.

Recursive's own engine

Recursive's own engine, rec, is tuned for Flint on newer Apple chips. The setup docs install it with one script, then launch a coding agent from your project folder:

curl -fsSL https://raw.githubusercontent.com/recursive-oss/rec/main/install.sh | bash
rec doctor
rec launch opencode

rec launch pi starts the Pi agent instead, which is Recursive's recommended pairing. rec serve starts an OpenAI-compatible server at http://127.0.0.1:11500/v1, so any tool that accepts a custom base URL can use it. On a Mac that doesn't qualify, rec refuses to download the model.

No hardware at all

Recursive's docs say a hosted version is free on OpenCode's free tier, and Socher's post also mentions Pi. If you go that route, remember that OpenCode had a Console breach this week that exposed some users' provider API keys. Don't paste keys you can't rotate.

A first prompt to try

Flint is built for repository work: reading unfamiliar code, tracing bugs, writing features and adding tests. Following Recursive's own advice, give it one small, explicit job:

Read src/forms/contact.ts. List every field that is not validated before submit. Then add validation for the email field only, and write one test for it.

If it handles that cleanly, try a bigger job split into two or three steps. If it wanders, narrow the instructions instead of switching models.

My read

I care less about the benchmarks than the download size. For 23 GB you get a coding assistant that works offline and never bills per token. We wrote last month about open models handling most AI traffic while the closed labs still collect most of the money, and Flint pushes further in that direction. A week of real use will tell you more about your own code than any launch chart.

For a small business, the draw is privacy. If you have client code or customer data that can't leave the building, you now have a local option worth testing, much like the small local models such as Bonsai 2 we covered in September. I wouldn't hand it the hardest problems. Keep a frontier model for those and let Flint take the routine work.

If you're trying to work out which parts of your workflow could run on a local model and which still need a paid one, that's the kind of question our free process audit is built to answer.