Model Export

Source: src/renderer/src/components/ModelExport.tsx

Overview

Export fine-tuned LoRA adapters as standalone models. Fuses the adapter weights with the base model and optionally quantizes the result.

Export Options

Precision	q_bits	Description
4-bit	4	Smallest size, some quality loss
8-bit	8	Balance of size and quality
Full precision	0	No quantization, largest size

Workflow

1. Open the Model Export page.
2. Select a fine-tuned adapter from the dropdown (only models with is_finetuned: true).
3. Choose a precision level.
4. Select an output directory via the file dialog.
5. Click Export.

The backend calls mlx_lm.fuse() to merge adapter weights into the base model. If quantization is selected (q_bits > 0), the fused model is also quantized.

API

List Adapters

GET /api/engine/models/adapters — returns models from the registry where is_finetuned is true.

Export

POST /api/engine/models/export

{
  "model_id": "adapter-uuid",
  "output_path": "/path/to/output",
  "q_bits": 4
}

Backend

Implementation: backend/app/engine/service.py method export_model().

fuse_kwargs = {"model": base_model, "adapter_path": adapter_path, "save_path": output_path}
if q_bits and q_bits > 0:
    fuse_kwargs["q_bits"] = q_bits
mlx_lm.fuse(**fuse_kwargs)