Flint INT4

Flint INT4

A 4-bit MLX copy of Recursive Flint for local use on Apple Silicon, about 22.6 GB, launched through Rec with pi or OpenCode. Suits Mac users and developers who want Flint behind a coding tool.

Made by RecursiveSI on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Flint INT4 is a 4-bit copy of Recursive Flint in the MLX format, made mainly for running locally on Apple Silicon. Projection and expert layers use group-64 affine INT4, and embeddings use INT8. The weights take up about 22.6 GB. For model details, the card sends you to Flint's own page.

Who it suits: Mac users who want to run Flint on their own machine, and developers who want it behind a coding tool. It also suits anyone who wants the smaller 4-bit weights rather than the full-precision release.

Industry | Who uses it | How they use it | Example
Software | Developers on Macs | Run a local model for coding help | Launch it through the Rec tool
AI tooling | Agent users | Connect a local model to pi or OpenCode | Start OpenCode with the Rec launcher
Everyday use | Privacy-minded users | Chat with a model that stays on the computer | Run Flint offline on an Apple Silicon Mac
Research | Quantization researchers | Study 4-bit affine weights with INT8 embeddings | Compare against the full-precision release

Anyone without an Apple Silicon Mac can skip the MLX files, and users of llama.cpp on Metal should take the separate Flint INT4 GGUF release instead.

Where to use it:

• Apple Silicon Macs, which the card names as the main target
• Rec, which launches the pi and OpenCode coding tools with this model
• Windows, macOS or Linux, if you only want to download the weights

Use case | Who | What they get back
Local text generation | Mac users | Replies generated on your own machine
Coding-tool backend | Developers | A local model behind pi or OpenCode through Rec
Weights download | Anyone | A local copy of the INT4 files in a folder you choose

How to set it up:

• Python with pip, if you only want to download the files
• A Mac with Apple Silicon and enough memory for about 22.6 GB of weights
• Rec, installed by following its own requirements and installation guide

1. Install the Hugging Face hub package.
2. Download the weights, either with the plain command or into a named folder.
3. Or install Rec and let it launch a tool that uses the model.

pip install -U huggingface_hub
hf download RecursiveSI/flint-int4

hf download RecursiveSI/flint-int4 --local-dir flint-int4

How to use it: After you install Rec, launch the tool you want, either pi or OpenCode. The card's other route is to download the weights on their own and load them with your MLX tooling.

rec launch pi
# Or use OpenCode:
rec launch opencode

Tool | What it does
rec launch pi | Starts the pi tool with the model
rec launch opencode | Starts OpenCode with the model
hf download | Saves the weights to a local folder

Example: You install Rec, run the pi launch command, and start working in the tool. Requests you type go to the local Flint INT4 model instead of a hosted service.

For newcomers and experts: If you are new, the easiest path is to install Rec, follow its guide, and run one launch command. You do not need to handle the weights yourself.

Experts will note the quantization split. Projections and experts are group-64 affine INT4, and embeddings are INT8, giving about 22.6 GB. Flint's own card holds the benchmarks, chat conventions and limits. Its training data mixed public and crawled web text, code, licensed data and synthetic data, followed by instruction tuning, reinforcement learning and preference data.

Install Rec following its guide
Run a launch command for pi or OpenCode
The tool loads the INT4 weights
You work with a local model

Good to know:

• Size: about 22.6 GB of INT4 and INT8 weights
• Target: mainly local inference on Apple Silicon
• Other format: llama.cpp on Metal needs the separate GGUF release
• License: Apache 2.0
• Cost to run: none named, no API key or account, only a Mac with enough memory
• Data it reaches: it downloads weights from Hugging Face and reads whatever you type into the tool
• Not for: the source names no forbidden uses, and the card points to Flint's own page for limitations

On Hugging Face, 13 November 2026

• Task: text generation
• Library: mlx
• Formats: Safetensors
• Downloads in the last 30 days: 22,586
• Likes: 30
• License: apache-2.0
• Quantized versions: 2
• Fine-tunes: 1
• Tags: flint, int4, safetensors, text-generation, conversational

Links

• Model: https://huggingface.co/RecursiveSI/flint-int4
• Files: https://huggingface.co/RecursiveSI/flint-int4/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download RecursiveSI/flint-int4

Then follow the usage in its model card. Nothing in this kit runs by itself.