- macOS and Linux:
~/.tapioca - Windows:
%USERPROFILE%\.tapioca
The current layout contains:
.tapioca/
├── models/ downloaded model weights and snapshots
├── adapters/ provider and imported LoRA weights with source metadata
├── recipes/ saved base-model and adapter combinations
├── voices/ reusable voice samples, transcripts, and metadata
├── runtime/ generated Python/MLX runtime environments
├── launch/ isolated coding-agent configuration
└── registry.json installed-model registry
Model directories use the resolved name and variant. For example:
~/.tapioca/models/glm-4.7-flash-q8_0/
~/.tapioca/models/flux2-klein-4b-q4-mlx/
Snapshot-based models contain the Hugging Face repository files. GGUF models
usually contain one .gguf weight file.
Set TAPIOCA_HOME before running Tapioca:
export TAPIOCA_HOME="/Volumes/ExternalSSD/tapioca"
tapioca listPowerShell:
$env:TAPIOCA_HOME = "D:\Tapioca"
tapioca listAdd the environment variable permanently using the operating system's normal environment settings if desired.
Release archives do not contain model weights. Leave enough free space for the download, runtime dependencies, temporary partial downloads, and outputs.
To reuse existing downloads or move weights between computers, follow Import and transfer. It explains how to:
- import a raw
.safetensorsLoRA withtapioca adapter import; - copy the complete managed adapter library with its metadata; and
- copy a catalog model directory and run
tapioca pull MODELto verify missing files and rebuild registration for the destination computer.
Do not copy registry.json by itself between computers. It stores absolute
paths from the machine that created it.
The detailed LoRA composition and storage layout is documented in LoRA adapters.