Local AI in 5MB. A native client for llama.cpp.
macOS installer download size
LM Studio ████████████████████████████ 556 MB
Ollama ████████ 160 MB
Jan 95 MB
pico ▏ 2 MB
Sources: lmstudio.ai/download, Ollama releases, Jan releases.
Every local LLM desktop app is an Electron build that ships its own copy of llama.cpp. The result is hundreds of megabytes of duplicated runtime and a Chromium tax for what is, structurally, a chat window and a download button.
pico assumes you already have llama-server on your machine (or installs it with one click via Homebrew), then puts a small native window on top. No bundled inference engine. No telemetry. No account.
Download the .app from releases, drag it to Applications, and open.
On first launch, pico checks for llama-server. If missing, it offers to run brew install llama.cpp for you.
- Discover. Search Hugging Face for GGUF models. Each file shows its quantization, size, and whether it fits in your RAM.
- Models. Click load on a downloaded model. pico spawns
llama-serverin the background and shows live load progress. - Chat. Stream messages with temperature and top-p controls. TTFT and tokens/sec are visible at the bottom.
When you eject a model, the underlying server process is killed.
- Tauri 2 shell, Rust backend, Svelte 5 frontend
- Backend calls
llama-server's OpenAI-compatible HTTP API on a local port - Models live in
~/Library/Caches/pico/models - Bundle is 3.3 MB because the inference engine is not bundled
No RAG, no MCP, no tool calling, no vision, no speculative decoding UI, no MLX backend yet. These are reasonable to want; they will arrive when they arrive. pico's scope is run a local model and chat with it, well.
Linux and Windows builds are planned for v0.2 via GitHub Actions.
git clone https://github.com/ddoemonn/pico
cd pico
pnpm install
pnpm tauri buildRequires Rust, Node 20+, pnpm.
MIT.