Flint NVFP4 GGUF

Flint NVFP4 GGUF

A 4-bit NVFP4 copy of the Recursive Flint text model as one GGUF file, built for NVIDIA GPUs through llama.cpp. For developers running a large local model with pi or OpenCode.

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 NVFP4 GGUF is a single GGUF file of the Recursive Flint text-generation model, stored in the NVFP4 4-bit format. The 22.9 GB file bundles the weights, the tokenizer and the chat template. The record lists 39.8B parameters.

Who it suits: Developers with a large NVIDIA GPU who want a big model running locally through llama.cpp.

Industry | Who uses it | How they use it | Example
Software | Developers | Run a local coding agent against their own server | pi or OpenCode pointed at a local llama-server
Research | AI labs | Test a large model on Blackwell-class hardware | A long-context run on one GPU
IT | Platform teams | Host a model on their own machines | An internal model endpoint on the team network

You can skip it if your GPU has under 48 GB of memory, or if you use a card that is not from NVIDIA.

Where to use it:

• llama.cpp with CUDA, through its llama-server program
• pi and OpenCode, two coding clients the card sets up for it
• Windows, macOS or Linux, with Python for the download step

Use case | Who | What they get back
Local model server | Developers | A llama-server that serves the model on your machine
Coding agent backend | Developers | pi or OpenCode talking to the local model
Long prompts | Researchers | A run at the full 262,144-token context

How to set it up:

• A recent llama.cpp release with Flint support, built with CUDA, and llama-server on your PATH
• An NVIDIA GPU with at least 48 GB of memory, or unified memory on a Spark machine
• The Hugging Face command line tool

1. Install llama.cpp with CUDA, following its build guide.
2. Install the Hugging Face CLI.
3. Download the model file and the two client config files.
4. Start llama-server with the model.

hf download RecursiveSI/flint-nvfp4-gguf flint-nvfp4.gguf pi-models.json opencode.json \
  --local-dir flint-nvfp4-gguf

llama-server -m flint-nvfp4-gguf/flint-nvfp4.gguf \
  --alias flint -ngl all -c 262144 -np 1 \
  -fa on -ctk f16 -ctv f16 --jinja \
  --temp 1 --top-p 0.95 --top-k 20 --min-p 0 \
  --host 127.0.0.1 --port 8080

How to use it: Keep llama-server running and open a second terminal in the folder that holds the downloaded model folder. Then start pi or OpenCode with the config that came with the download.

Tool | What it does
llama-server | Serves the model from the GGUF file
pi | A coding client, started with the pi-models.json config
OpenCode | A coding client, started with the opencode.json config

mkdir -p .pi-flint
cp flint-nvfp4-gguf/pi-models.json .pi-flint/models.json
PI_CODING_AGENT_DIR="$PWD/.pi-flint" pi --provider flint-local --model flint

OPENCODE_CONFIG="$PWD/flint-nvfp4-gguf/opencode.json" opencode --standalone

Example: With the server running, you start OpenCode and ask it to explain a function in your project. The request goes to the local Flint model, and the answer comes back in your terminal.

For newcomers and experts: If you are new, this is a heavy download that needs a big GPU, so read the card and check your memory before you start. Run the server first, then one client.

If you know the field, the card says the file holds packed NVFP4 weights and scales. Run with the jinja flag and keep reasoning_content when you replay a conversation. The client configs set a 262,144-token window and a 65,536-token generation budget, so match them if you change the context flag.

Download the GGUF file and configs
Start llama-server on the GPU
Open pi or OpenCode
Ask questions against the local model

Good to know:

• License: Apache 2.0
• Cost to run: your own NVIDIA GPU, with at least 48 GB of memory advised
• Data it reaches: the files you download, and prompts you send to your own local server
• Tested on: four NVIDIA setups, among them the B300 and the DGX Spark

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: 132,757
• Likes: 40
• License: apache-2.0
• Fine-tunes: 1
• Tags: gguf, nvfp4, text-generation, conversational

Links

• Model: https://huggingface.co/RecursiveSI/flint-nvfp4-gguf
• Files: https://huggingface.co/RecursiveSI/flint-nvfp4-gguf/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download RecursiveSI/flint-nvfp4-gguf

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