01What Flint Local Coding Model Is and How It Works

Flint local coding model is an AI coding tool that runs entirely on your own computer. Unlike most AI coding tools that send your code to a remote server, Flint keeps everything local. Every prompt, file, and line of code stays on your machine.

This means your proprietary logic never leaves your control.

Flint is a 40 billion parameter model. But it does not use all 40 billion parameters at once. It uses a method called mixture of experts.

Think of it like a team of specialists. When a task comes in, only the right specialists activate. In Flint's case, roughly 2.5 billion parameters fire per token.

This gives you the quality of a large model but at a fraction of the compute cost. That is what makes it fast without being a weak model.

02Key Features and What's New

Flint is not just a code completion tool. It works like an agent. It can inspect a full repository, find a bug, write a fix, add a regression test, and check if the fix passes.

It handles multiple files, multiple steps, and multiple turns.

  • Local inference: Everything runs on your machine, no remote server.
  • Open weights under Apache 2.0: Anyone can download and use it for free.
  • Fast speed: Over 125 tokens per second on an M4 Pro MacBook Pro. Over 150 tokens per second on an NVIDIA RTX Spark.
  • Day-one ecosystem support: VLLM and OpenHands announced support on release day.
  • Built by Recursive in under four months using the Eureka Machine, an automated research system.

03How to Set Up Flint on Your Machine

  1. Pick your hardware. You need a machine with a capable GPU, like an NVIDIA RTX Spark or an M4 Pro MacBook Pro. Check the file sizes: the original BF16 checkpoint is about 80GB, the CPU quantized version is about 22.9GB.
  2. Download the model weights from Hugging Face at RecursiveSI/Flint. Choose the checkpoint that matches your hardware and storage.
  3. Set up a local inference server. The cleanest way is VLLM. They have published Docker launch instructions and their kernels are optimized for Flint's architecture.
  4. If you prefer Ollama, check their registry for Flint variants and follow their install instructions.
  5. For agentic coding workflows, install OpenHands. Point it at your local Flint instance and connect it to a repository. The agent will start working.
  6. Test your setup with a simple coding task to make sure everything runs correctly.

04How to Use Flint for Real Coding Tasks

Flint is designed to work like an agent. You give it a repository, and it can navigate the code, find a bug, write a fix, add a regression test, and check if the fix passes. It handles multiple files and multiple steps.

For example, you point Flint at a GitHub issue. Flint scans the repository, identifies the problem, writes the code to fix it, creates a test to make sure the bug does not come back, runs the test, and reports the result. All of this happens on your local machine.

In the episode, Julian showed that Flint can handle these tasks with multiple turns. The agent loop includes editing files, running commands, reading test results, and trying again if needed. This is the same workflow used by professional developers.

05Benchmark Results and Tests

Flint has been tested on several benchmarks. Here are the published numbers from Recursive:

  • SWE-bench Verified: 74.2%. This tests whether a model can fix real GitHub issues.
  • SWE-bench Pro V2: 89.7%. This is a harder version of the same test.
  • Terminal Bench 2.1: 54.9%. This focuses on agentic terminal tasks.

These numbers are from Recursive's own evaluation setup. Benchmarks can vary by harness and conditions, but these results are strong for a model you run locally.

On hardware, Recursive reported over 125 tokens per second on an M4 Pro MacBook Pro and over 150 tokens per second on an NVIDIA RTX Spark. VLLM also reported around 500 tokens per second on a single GB200 GPU, including on a 100,000 token context.

06Tips, Limits, and Mistakes to Avoid

Be careful with hardware requirements. The original BF16 checkpoint is about 80GB. That means you need serious storage and a capable machine.

The CPU quantized version is about 22.9GB, but the GPU memory requirements are still higher than the file size suggests. Do not assume any laptop will handle it comfortably.

Benchmark numbers are from Recursive's test conditions. Your results may differ based on your hardware and setup. The numbers are a good guide, not a guarantee.

If you use Ollama, make sure you pick the right variant for your machine. Not all Flint variants will run well on every GPU. Check the model size and your available memory before downloading.

For agentic workflows, start with a simple repository. Let Flint get used to your codebase before giving it a complex multi-file bug. Always review the changes Flint makes.

Even a good model can make mistakes.

07Who Flint Is For and Why It Matters

Flint is for developers and businesses who want to keep their code private. If you work on proprietary software or sensitive logic, sending your code to a remote server is risky. Flint runs everything on your machine, so your code never leaves.

It is also for anyone who wants fast local AI coding. You get the quality of a large model but with the speed of a small one. That is useful for teams that need quick feedback without relying on cloud APIs.

Business owners who set up Flint now have an advantage. Most have not figured out local AI coding models. They have not set up Flint or run a coding agent locally.

The first people on this will have a head start.

Flint is open weights and free. You can download it, run it, and use it without paying per prompt or per file. No subscription required.