Install
Get the app
curl -fsSL https://raw.githubusercontent.com/kyndlo/tapioca/main/scripts/install.sh | shmacOS Apple Silicon. Windows users can grab the x64 or ARM64 bundle from Releases.
Local AI, minus the drama
Your models.
Ready to roll.
Run language, image, video, and coding-agent models on your own Mac or Windows PC—with one small, friendly command-line tool.
✓ Open source✓ Local-first✓ No account required
GLM
MLX
GGUF
tapioca
› tapioca run qwen3:8b-q4_k_mmodel ready · 6.2 GB · metalyou › Build me something delightful.tapioca › Let's cook. ▋01 · Beginner lane
From zero to chatting
in three tiny steps.
No model archaeology degree required. Tapioca downloads what you need and remembers where it lives.
Choose
Pick a model
tapioca catalogSee download size, memory guidance, GPU needs, and the right variant for your computer.
Run
Say hello
tapioca run qwen3:8b-q4_k_mIf the model is missing, Tapioca pulls it automatically. Type /bye when you're done.
Not sure which model? Use the friendly model chooser →
02 · Bring your computer
One CLI. Three hardware worlds.
Apple Silicon
Metal, MLX, MFLUX, and native llama.cpp. A particularly cozy home for local models.
Windows x64
Vulkan text models, NVIDIA CUDA, plus DirectML image generation for AMD and Intel.
Windows ARM64
Native ARM64 Tapioca, CPU llama.cpp, and ONNX image generation without x64 emulation.
03 · Six verbs, lots of power
A pleasantly small command surface.
Learn the shape once, then move from a chat to an API, an image, or a full coding agent.
›
pull
Download a model once
tapioca pull qwen3:8b-q4_k_m›
run
Chat in your terminal
tapioca run qwen3:8b-q4_k_m›
serve
Open an API for your apps
tapioca serve qwen3:8b-q4_k_m›
image
Create images locally
tapioca image sd-turbo --prompt "A pearl astronaut"›
video
Turn prompts into motion
tapioca video ltx-video:2b-fp16 --prompt "Clouds rolling in"›
launch
Power your coding agent
tapioca launch opencode qwen3-coder:30b-mlx04 · The local studio
Words in.
Pixels and motion out.
Tapioca understands platform-native diffusion: MLX and MFLUX on Mac, CUDA and DirectML on Windows, and native ONNX on Windows ARM64.
IMAGE RECIPEfox-studio
base sd-turbo:onnx-directmlprompt “a red fox in snow”size 512 × 512steps 4tapioca image fox-studio05 · Expert lane
Simple outside.
Serious underneath.
Use Tapioca as an orchestration layer, a local OpenAI-compatible endpoint, or a portable runtime for your team.
Your tools Codex · Claude Code · OpenCode
OpenAI-compatible API
Tapiocacatalog · routing · lifecyclellama.cppGGUFMLX / MFLUXApple SiliconDiffusersCUDAONNXDirectML / ARM
OpenAI-compatible API
Use /v1/chat/completions with tools, streaming, and familiar clients.
Backend-aware catalog
A model variant declares its runtime, memory guidance, GPU needs, and platform.
Composable models
Save reusable recipes that combine a base model, LoRAs, adapters, and presets.
Local by design
Models and generated media stay on your machine under ~/.tapioca.
Your machine has been waiting
Ready to roll?
Start small. Pull a model. Make something weird and wonderful.