What is ForgeAI?
ForgeAI is a local-first desktop application for working with AI models. Load GGUF and SafeTensors models, inspect their architecture, quantize for deployment, fine-tune with LoRA/QLoRA, merge multiple models into hybrids, explore datasets, and test inference — all without uploading anything to the cloud.
Load & Inspect
Import models from disk. Analyze architecture, memory layout, tensor composition, capabilities, and runtime compatibility with 3D visualization.
Train & Fine-Tune
Fine-tune models with LoRA, QLoRA, SFT, DPO, or Full training. Target specific capabilities like reasoning, code, or math. Perform layer surgery.
Merge Models
Combine 2–5 parent models using 12 methods — SLERP, TIES, DARE, DeLLa, Frankenmerge, MoE conversion, and more.
Explore Data
Load and analyze datasets from local files or HuggingFace. Support for JSON, JSONL, CSV, and Parquet formats.
Capabilities
| Module | Code | Description |
|---|---|---|
| Dashboard | 00 | System overview with real-time module activity tracking |
| Load | 01 | Import GGUF files, SafeTensors files, or sharded model folders |
| Inspect | 02 | 3D architecture viz, memory layout, capabilities, runtime compatibility |
| Compress | 03 | Quantize GGUF models across 7 levels (Q2_K through Q8_0) |
| Hub | 04 | Download models from HuggingFace, manage local model library |
| DataStudio | 10 | Explore datasets (JSON/JSONL/CSV/Parquet) from local or HuggingFace |
| Training | 06 | Fine-tune (LoRA/QLoRA/SFT/DPO/Full) and layer surgery |
| Convert | 05 | Convert SafeTensors → GGUF with configurable output types |
| M-DNA Forge | 08 | Merge 2–5 models with 12 methods, 3D visualization, presets |
| Test | 09 | Run inference with real-time token streaming and performance stats |
| Settings | 07 | Theme, fonts, GPU detection, 3 environment managers |
Tech Stack
Frontend
SvelteKit 5 with Svelte 5 runes (
$state, $derived, $effect)Backend
Rust via Tauri v2 — native desktop on Linux, macOS, Windows
Tensors
Candle (Rust ML framework) for merge tensor math
GGUF
llama.cpp for quantization and inference
Training
PyTorch + PEFT + TRL via managed Python subprocess
Datasets
Apache Arrow + Parquet for native Rust dataset parsing
Supported Formats
Model Formats
| Format | Load | Inspect | Compress | Convert | Merge | Train | Surgery | Test |
|---|---|---|---|---|---|---|---|---|
| GGUF | ✓ | ✓ | ✓ | Output | ✓ | ✓ | ✓ | ✓ |
| SafeTensors | ✓ | ✓ | — | Input | ✓ | ✓ | ✓ | ✓ |
| Sharded Folders | ✓ | ✓ | — | Input | ✓ | ✓ | ✓ | ✓ |
Dataset Formats
| Format | DataStudio | Training | HuggingFace |
|---|---|---|---|
| JSON | ✓ | ✓ | ✓ |
| JSONL | ✓ | ✓ | ✓ |
| CSV | ✓ | ✓ | ✓ |
| Parquet | ✓ | ✓ | ✓ |
Navigation
ForgeAI uses a sidebar with modules organized into groups:| Group | Modules |
|---|---|
| MODEL | 01 Load · 02 Inspect · 03 Compress |
| DATA | 04 Hub · 10 DataStudio · 06 Training |
| TOOLS | 05 Convert · 08 M-DNA · 09 Test |
| SYSTEM | 00 Dashboard · 07 Settings |