Load

Import AI models from disk — GGUF files, SafeTensors files, or sharded model folders

Load (01)

The Load module imports AI models into ForgeAI. It parses model headers without reading the full file, extracting metadata, tensor maps, and architecture information in under a second.

Load Options

GGUF File

Single .gguf file — quantized models from the llama.cpp ecosystem

SafeTensors File

Single .safetensors file — HuggingFace model weights

Model Folder

Directory with .safetensors + config.json — sharded HuggingFace models

What Gets Parsed

On load, ForgeAI reads the model header and extracts:

• Architecture (e.g., LlamaForCausalLM, MistralForCausalLM, Qwen2ForCausalLM)
• Layer count and context length
• Embedding dimensions
• Parameter count (computed from tensor shapes)
• Quantization type (GGUF only)
• Tensor map — names, dtypes, and shapes of all tensors
• Raw metadata — all key-value pairs from the model header

For folder-based models, it additionally detects:

• Number of shards
• Presence of config.json
• Presence of tokenizer files

Hero Panel

Field	Values
STATUS	IDLE, LOADING, LOADED, ERROR
FORMAT	GGUF or SAFETENSORS
SIZE	File size on disk
PARAMS	Parameter count (e.g., 7.24B)
TENSORS	Total tensor count
QUANT	Quantization type (e.g., Q4_K_M)

Metadata Grid

After loading, a grid displays key model properties:

Field	Description
ARCHITECTURE	Model family (LlamaForCausalLM, etc.)
LAYERS	Number of transformer layers
CONTEXT	Maximum sequence length
EMBEDDING	Hidden dimension size
QUANT	Quantization type
TENSORS	Total tensor count
PARAMS	Parameter count
SIZE	File size on disk

Tensor Map Preview

A table showing tensor names, dtypes, and shapes. For large models the preview is capped with a count of remaining tensors.

Raw Metadata

A scrollable list of the first 20 metadata key-value pairs from the model header.

Workflow

1. Click a load button

   Choose LOAD GGUF FILE, LOAD SAFETENSORS FILE, or LOAD MODEL FOLDER

2. Select file or folder

   Pick your model in the system dialog

3. Review metadata

   Header parsing takes under 1 second. Review model metadata, tensor map, and raw metadata.

4. Use the model

   The model is now available in Inspect, Compress, Convert, Training, M-DNA, and Test modules. Click UNLOAD to clear it.

Loading only parses the header — it does not load the full model into memory, so even 70B+ models load instantly.