flint-GGUF

flint-GGUF

Runs Flint, a 39.8B mixture-of-experts model for coding agents that uses about 2.5B parameters per token, from Unsloth GGUF files. For developers who run coding agents locally.

Made by unsloth 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-GGUF holds Unsloth GGUF versions of Flint, a language model from Recursive built for coding agents. It has 39.8 billion parameters in total, and about 2.5 billion are active for each token.

Flint is meant to explore codebases, fix bugs, add features and write tests. It reads and writes text, reasons, and supports multi-turn tool use.

Who it suits: Developers who want to run a coding agent model on their own machine, and teams that compare local models on software engineering tasks.

Industry | Who uses it | How they use it | Example
Software | Developers using coding agents | Run a local model behind an agent | Have it fix a bug in a repository
Engineering teams | Teams that maintain large codebases | Let an agent write tests | Generate tests for a new feature
Research | People comparing coding models | Score it on software engineering tasks | Compare it with other local models

Skip it if you need image or audio input, since Flint takes text in and gives text out.

Where to use it:

• A local runner that opens GGUF files
• Unsloth Desktop, which the card shows with thinking controls
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Bug fixing | Developers | Code changes proposed after the model investigates a codebase
Feature work | Agent builders | Implementations produced over several tool-using turns
Test writing | Maintainers | New tests for existing code

How to set it up:

• Python and pip to run the download commands
• Enough memory for the files, since the 4-bit weights are about 19.9 GB by Hugging Face estimate
• A GGUF-capable runner, or the Unsloth guide for running Flint

1. Install the Hugging Face tools.
2. Download the GGUF files from this repository.
3. Open them in your runner, following the Unsloth guide for Flint.

pip install -U huggingface_hub
hf download unsloth/flint-GGUF

How to use it: Load the files in your runner and connect an agent or chat to it. The authors ran their evaluations with a temperature of 1.0, a top-p of 0.95 and a top-k of 20.

Property | Flint
Total parameters | 39.8B
Active per token | About 2.5B
Layers | 49 in all, the first 2 dense and the other 47 mixture-of-experts
Attention | 35 layers look at a local window and 14 layers attend to the full context
Experts | 256 routed, 6 chosen per token, plus one shared expert that is always on
Context | 262,144 tokens in the documented setup

Example: You point a coding agent at a repository with a failing test and ask it to find and fix the bug. The model investigates the code over several turns, calls tools and proposes a fix.

For newcomers and experts: If you are new to coding agents, start by loading the files and asking one plain coding question. Move to an agent setup once the basic chat works.

If you know the field, it is a decoder-only mixture-of-experts transformer with sliding-window attention on most layers. Each MoE layer routes a token to 6 of 256 experts and also runs one shared expert.

Download the GGUF files
Open them in a local runner
Connect a coding agent
Model investigates, edits and tests

Good to know:

• The authors measured the scores and ran Terminal-Bench with Terminus 2 and SWE-bench with mini-swe-agent
• Their Terminal-Bench score counts solutions found through web retrieval as failures
• The original repository holds the BF16 checkpoint, and quantized variants are prepared separately
• Hugging Face lists 1 quantized and 1 fine-tuned version made by other people

Title: Release evaluation scores by the authors
Terminal-Bench 2.1: 54.9 percent
SWE-bench Pro v2: 89.7 percent
SWE-bench Verified: 74.2 percent

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the files you download and the code and prompts you give it
• Not for: input other than text

On Hugging Face, 13 November 2026

• Task: text generation
• Parameters: 39.8B
• Weights: about 19.9 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 325,104
• Likes: 376
• License: apache-2.0
• Based on: RecursiveSI/flint
• Quantized versions: 1
• Fine-tunes: 1
• Tags: gguf, flint, unsloth, text-generation, imatrix, conversational

Links

• Model: https://huggingface.co/unsloth/flint-GGUF
• Files: https://huggingface.co/unsloth/flint-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download unsloth/flint-GGUF

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