Your local models, ready to roll.
Tapioca is a local AI runtime and desktop studio for language, speech, audio, image, and video models. It downloads supported models on demand, chooses a native backend for the computer, and can expose local models to coding agents.
Use the friendly desktop app or the full CLI—both run the same private, local-first engine.
Tapioca Desktop brings the complete local workflow into one interface:
- Chat with installed models and see model reasoning as it streams.
- Generate and manage images, videos, music, sound effects, speech, and cloned voices.
- Record voice-cloning references with a live microphone level meter.
- Choose guided image resolutions, video duration, orientation, quality, and workload presets while retaining exact expert controls.
- Pull LoRAs from Hugging Face, Civitai, or ModelScope, import local
.safetensorsfiles, and assign them with model-family compatibility guidance. - Pull models, inspect hardware requirements, and track local storage.
- Launch Codex, Claude Code, OpenCode, OpenClaw, and Hermes with a local model.
- Navigate with a consistent accessible icon system and review long agent paths without layout overflow.
- Monitor active jobs and keep generated media in a local gallery.
Download the latest installer from GitHub Releases:
| Platform | Desktop package | CLI package |
|---|---|---|
| macOS Apple Silicon | .dmg |
tapioca-darwin-arm64.tar.gz |
| Windows x64 | .exe |
tapioca-windows-amd64.zip |
| Windows ARM64 | .exe |
tapioca-windows-arm64.zip |
| Linux x64 | .AppImage |
tapioca-linux-amd64.tar.gz |
| Linux ARM64 | .AppImage |
tapioca-linux-arm64.tar.gz |
- Download the desktop app or follow the complete beginner guide
- Run your first local model
- Choose a model for your computer
- Explore the task-specific guides below.
tapioca catalog
tapioca catalog update
tapioca update [--check]
tapioca run gemma3:12b-mlxTapioca pulls a catalog model automatically when it is first used. Enter
/bye or press Ctrl-D to leave an interactive chat and stop its server.
- Text chat and local APIs
- LLM and agent operating guide
- Agent API reference
- Install the Codex plugin or Claude Code skill
- Coding agents: Codex, Claude Code, OpenCode, OpenClaw, and Hermes
- Image generation
- Video generation
- Music and sound generation
- Text to speech and voice cloning
- LoRA adapters and model composition
- Replaceable video engines
- Command reference
- Files, models, and disk storage
- Import existing LoRAs and transfer model files
- Troubleshooting
- Build from source
Agents can start at tapioca.rootfruit.cc/llm or read the machine-oriented llms-full.txt.
| Capability | macOS Apple Silicon | Windows x64 | Windows ARM64 | Linux x64 | Linux ARM64 |
|---|---|---|---|---|---|
| GGUF text models | Metal llama.cpp | Vulkan llama.cpp | CPU llama.cpp | Vulkan llama.cpp | Vulkan llama.cpp |
| Native MLX text | Yes | No | No | No | No |
| Speech and voice cloning | MLX or MPS/CPU | CUDA or CPU | CPU | CUDA or CPU | CPU |
| Music and sound generation | Stable Audio 3 | Stable Audio 3 CUDA | CPU fallback | Stable Audio 3 CUDA | CPU fallback |
| Image generation | MLX/MFLUX | CUDA or ONNX DirectML | ONNX Runtime CPU | NVIDIA CUDA | Not yet |
| Video generation | MLX | NVIDIA CUDA/Diffusers | Not yet | NVIDIA CUDA/Diffusers | Not yet |
| Coding-agent APIs | Yes | Yes | Yes | Yes | Yes |
| Desktop app | Yes | Yes | Yes | Yes | Yes |
Run tapioca catalog for the exact models, download sizes, memory guidance,
GPU requirements, and supported platforms in the installed release.
Run tapioca catalog update to fetch the checksummed catalog independently of
the binary, or use Refresh catalog in Desktop settings. The compiled
catalog remains an offline fallback. tapioca update verifies the latest
GitHub Release checksum and replaces the CLI/runtime without touching models,
adapters, voices, or outputs.
MiniMax-H3 is available for Apple Silicon and Windows/Linux NVIDIA as
minimax-h3. It supports text-to-video, image-to-video, and native stereo
audio through a managed ComfyUI runtime. The four-file model bundle is about
41 GiB; Apple Silicon needs at least 48 GiB unified memory, while the CUDA
variant is designed for 16 GiB GPUs such as the RTX 4070 Ti SUPER. Compatible
MiniMax-H3 transformer LoRAs can be stacked with repeated --adapter flags.
On Windows x64, MiniMax-H3 automatically provisions its pinned Python and
CUDA-enabled media runtime; users need a current NVIDIA driver, not Git, a
system Python installation, or the CUDA Toolkit.
Qwen3.8-27B is available as qwen3.8. Apple Silicon uses the optimized
27b-mlx snapshot by default, while Windows and Linux use the portable
27b-q4_k_m GGUF variant. It is aimed at agentic coding, reliable tool use,
and long-horizon reasoning; use 27b-q5_k_m when additional quality is worth
the larger download and memory footprint.
Krea 2 Turbo is available as krea-2-turbo for Windows/Linux NVIDIA and,
experimentally, Apple Silicon Macs with large unified memory. The curated
Diffusers snapshot is about 34 GiB and defaults to 8 steps at 1024×1024. It is
a gated model: users must accept its Hugging Face terms, set HF_TOKEN, review
the Krea 2 Community License, and run
tapioca pull krea-2-turbo --accept-license. Tapioca does not accept provider
terms or store access tokens on a user's behalf. --accept-license records the
local Tapioca acknowledgement; it does not approve access in the user's
Hugging Face account.
Qwen Image 2.1 is available as qwen-image-2.1 for Windows/Linux NVIDIA with
Diffusers 0.41. It supports image editing, native RGBA transparency, improved
text rendering, and multiple references. The roughly 31 GiB BF16 snapshot is
public but restricted to non-commercial research/evaluation by the Qwen
Research License, so Tapioca requires --accept-license without requiring an
Hugging Face token.
tapioca catalog
tapioca catalog update
tapioca update [--check]
tapioca list
tapioca pull MODEL[:VARIANT] [--accept-license]
tapioca run MODEL [--context TOKENS]
tapioca serve MODEL [--port 11435] [--context TOKENS]
tapioca image MODEL --prompt TEXT [--seed NUMBER | --random-seed] [--output image.png]
tapioca edit MODEL --image FILE [--image FILE] --prompt TEXT [--seed NUMBER | --random-seed]
tapioca video MODEL --prompt TEXT [--image start.png] [--seconds N | --frames N] [--seed NUMBER | --random-seed] [--output video.mp4]
tapioca audio MODEL --prompt TEXT [--seconds N] [--seed NUMBER | --random-seed] [--output audio.wav]
tapioca tts MODEL --text TEXT [--voice NAME] [--output speech.wav]
tapioca voice (create|list|inspect|remove) [NAME]
tapioca adapter (inspect|pull|import|list) [REFERENCE]
tapioca create NAME --base MODEL [--adapter REFERENCE]
tapioca launch CLIENT MODEL [-- CLIENT_ARGS...]
See the command reference for flags and examples.
Use --seconds when you know the approximate clip length but not the valid
frame count for a model. Tapioca selects and prints the nearest supported frame
count. Do not combine --seconds with --frames:
tapioca video minimax-h3 \
--prompt 'A presenter says exactly: "Hello from Tapioca."' \
--seconds 5 --random-seed --output hello.mp4Image, edit, and video commands use seed 0 by default. --random-seed
generates and prints a portable random seed. Copy that value into
--seed NUMBER to reproduce the same generation settings later. The two seed
flags cannot be combined.
The built-in catalog provides tested defaults, while image and video commands
can also accept direct hf://OWNER/REPOSITORY base-model references. LoRA
support is backend-specific: MFLUX image models, CUDA Diffusers pipelines,
Wan models through MLX-video, and MiniMax-H3 through Tapioca's managed video
engine can load compatible adapters. Tapioca validates
obvious model-family mismatches, but the upstream runtime remains the final
authority for less common adapters. Windows x64 AMD/Intel GPUs use ONNX
DirectML; Windows ARM64 uses native ONNX Runtime CPU inference.
Tapioca's original code, documentation, and project assets are open source under the Apache License 2.0. It permits commercial and private use, modification, and redistribution while preserving copyright, license, and notice requirements, and it includes an explicit patent grant.
Bundled runtimes, installed dependencies, models, and LoRAs retain their own licenses. See NOTICE and third-party software and model notices before redistributing a binary bundle or model files.
