Load → Inspect → Test
1
Load a model
- Click Load (01) in the sidebar
- Click LOAD GGUF FILE (or LOAD SAFETENSORS FILE / LOAD MODEL FOLDER)
- Select your model file in the dialog
- Wait for header parsing (< 1 second)
2
Inspect the architecture
- Click Inspect (02) in the sidebar
- Explore the isometric 3D visualization — hover layers for details
- Review memory distribution, quantization breakdown, and runtime compatibility
- Check capability detection (reasoning, code, math, etc.)
3
Run inference
- Click Test (09) in the sidebar
- Click USE LOADED to select your model
- Type a prompt or click a quick test preset (CODE, MATH, REASON, CREATIVE, INSTRUCT, CHAT)
- Configure GPU layers, temperature, and max tokens
- Click GENERATE and watch tokens stream in real time
For GGUF inference, llama.cpp tools must be installed. Go to Settings (07) → llama.cpp Tools → DOWNLOAD & INSTALL.
Download → Quantize
1
Download from HuggingFace
- Click Hub (04) in the sidebar
- Enter a repo ID (e.g.,
TheBloke/Mistral-7B-Instruct-v0.2-GGUF) - Click FETCH, then DOWNLOAD on a GGUF file
2
Load and quantize
- In Library view, click LOAD on the downloaded model
- Go to Compress (03)
- Pick a quantization preset (MOBILE, BALANCED, or QUALITY)
- Click QUANTIZE MODEL and choose output path
Explore a Dataset
1
Load from HuggingFace
- Click DataStudio (10) in the sidebar
- Switch to HUGGINGFACE source
- Enter a dataset repo ID (e.g.,
tatsu-lab/alpaca) - Click FETCH to see available files
- Click DOWNLOAD on a Parquet or JSON file
- Dataset auto-loads after download
2
Analyze
Review metadata (rows, columns, format), column analysis (dtypes, null counts), and preview data in the table view.
Fine-Tune a Model
1
Select model and dataset
- Click Training (06) in the sidebar
- Browse for a model (GGUF or SafeTensors)
- Browse for a dataset (JSON, JSONL, CSV, or Parquet)
2
Choose method and preset
- Select a training method (LoRA, QLoRA, SFT, DPO, Full)
- Pick a VRAM preset (LOW VRAM ~4GB, BALANCED ~6GB, QUALITY ~12GB, MAX QUALITY ~24GB)
- Optionally target specific capabilities (reasoning, code, math, etc.)
3
Train
Click START TRAINING, monitor real-time progress (loss, epoch, VRAM usage, ETA).
Merge Two Models
1
Load parents
- Click M-DNA (08) in the sidebar
- In the Files tab, click + LOAD FILE twice to load 2 models
2
Configure merge
- Go to Settings tab, pick a preset (QUICK BLEND, SMOOTH MERGE, TASK TUNER, etc.)
- Or manually select a method and tune parameters
- Set output format (SafeTensors or GGUF) and output path
3
Build
- Click BUILD MERGE
- Monitor progress in the sidebar and status bar
- Result shows output path and size