YaGGUF - Yet Another GGUF Converter

There are simultaneously too many and not enough GGUF converters in the world.

More screenshots

Features

• llama.cpp under the hood - so that part works
• Download - automatically download models and their auxiliary files from HuggingFace
• Convert - safetensors and PyTorch models to GGUF format
• Quantize - to multiple formats at once
• Cross-platform - works on Windows and Linux (and probably Mac but untested)
• Easy - auto-installs an environment + llama.cpp + CPU binaries for quantizing
• Flexible - can use any local llama.cpp repo or binary installation for quantizing
• Minimal mess - virtual environment prevents conflicts with your python setup

Advanced Features

• Single or split files mode - Generate single or split files for intermediates and quants
• Split/Merge Shards - Split, merge, or resplit GGUF and safetensors files with custom shard sizes
• Importance Matrix - Generate or reuse imatrix files for better low-bit quantization (IQ2, IQ3)
• Imatrix Statistics - Analyze importance matrix files to view statistics
• Custom intermediates - Use existing GGUF files as intermediates for quantization
• Enhanced dtype detection - Detects model precision (BF16, F16, etc.) from configs and safetensors headers
• Model quirks detection - Handles Mistral format, pre-quantized models, and architecture-specific flags
• Vision/Multimodal models - Automatic detection and two-step conversion (text model + mmproj-*.gguf)
• Sentence-transformers - Auto-detect and include dense modules for embedding models
• VRAM Calculator - Estimate VRAM usage and recommended GPU layers (-ngl) for GGUF models

Quantization Types

All quantization types from llama.cpp are supported. Choose based on your size/quality tradeoff:

| Type | Size | Quality | Category | Imatrix | Notes | |------|------|---------|----------|---------|-------| | F32 | Largest | Original | Unquantized | - | Full 32-bit precision | | F16 | Large | Near-original | Unquantized | - | Half precision | | BF16 | Large | Near-original | Unquantized | - | Brain float 16-bit | | Q8_0 | Very Large | Excellent | Legacy | - | Near-original quality | | Q5_1, Q5_0 | Medium | Good | Legacy | - | Legacy 5-bit | | Q4_1, Q4_0 | Small | Fair | Legacy | - | Legacy 4-bit | | Q6_K | Large | Very High | K-Quant | Suggested | Near-F16 quality | | Q5_K_M | Medium | Better | K-Quant | Suggested | Higher quality | | Q5_K_S | Medium | Better | K-Quant | Suggested | 5-bit K small | | Q4_K_M | Small | Good | K-Quant | Suggested | 4-bit K medium | | Q4_K_S | Small | Good | K-Quant | Suggested | 4-bit K small | | Q3_K_L | Very Small | Fair | K-Quant | Recommended | 3-bit K large | | Q3_K_M | Very Small | Fair | K-Quant | Recommended | 3-bit K medium | | Q3_K_S | Very Small | Fair | K-Quant | Recommended | 3-bit K small | | Q2_K | Tiny | Minimal | K-Quant | Recommended | 2-bit K | | Q2_K_S | Tiny | Minimal | K-Quant | Recommended | 2-bit K small | | IQ4_NL | Small | Good | I-Quant | Recommended | 4-bit non-linear | | IQ4_XS | Small | Good | I-Quant | Recommended | 4-bit extra-small | | IQ3_M | Very Small | Fair | I-Quant | Recommended | 3-bit medium | | IQ3_S | Very Small | Fair+ | I-Quant | Recommended | 3.4-bit small | | IQ3_XS | Very Small | Fair | I-Quant | Required | 3-bit extra-small | | IQ3_XXS | Very Small | Fair | I-Quant | Required | 3-bit extra-extra-small | | IQ2_M | Tiny | Minimal | I-Quant | Required | 2-bit medium | | IQ2_S | Tiny | Minimal | I-Quant | Required | 2-bit small | | IQ2_XS | Tiny | Minimal | I-Quant | Required | 2-bit extra-small | | IQ2_XXS | Tiny | Minimal | I-Quant | Required | 2-bit extra-extra-small | | IQ1_M | Extreme | Poor | I-Quant | Required | 1-bit medium | | IQ1_S | Extreme | Poor | I-Quant | Required | 1-bit small |

Quick Guide:

• Bigger is better (more precision)
• For best quality use F16 or Q8_0
• For decent quality use Q6_K or Q5_K_M
• Medium quality... Use Q4_K_M
• For smallest size use IQ3_M or IQ2_M with importance matrix

Requirements

• Python 3.8 or higher
• Git 2.20 or higher (if you want the update tab to work)

Installation - Windows

# Clone the repository
    git clone https://github.com/usrname0/YaGGUF.git
    cd YaGGUF
# Run the launcher script for Windows (runs a setup script if no venv detected):
    .\run_gui.bat

Installation - Linux

# If you want to select folders via the gui install tkinter (optional):
    sudo apt install python3-tk      # Ubuntu/Debian
    sudo dnf install python3-tkinter # Fedora/RHEL
    sudo pacman -S tk                # Arch

# Clone the repository
    git clone https://github.com/usrname0/YaGGUF.git
    cd YaGGUF

# Run the launcher script for Linux (runs a setup script if no venv detected):
    ./run_gui.sh

Usage

Windows:

• Double-click .\run_gui.bat

Linux:

• Use terminal ./run_gui.sh

The GUI will automatically open in your browser on a free port like: http://localhost:8501

License

MIT License - see LICENSE file for details

Credits

• llama.cpp - GGUF format and conversion/quantization tools
• HuggingFace - Model hosting and transformers library
• Streamlit - Pythonic data apps