Private AI chat desktop application with local LLM support.
All inference happens on your machine — no cloud, no data sharing.

• What is this?
• Demo
• Key Features
• Installation & Setup
• How to Start Using
• System Requirements
• Supported Models
• Privacy and Security
• Acknowledgments
• License

Oxide Lab is a native desktop application for running large language models locally. Built with Rust and Tauri v2, it provides a fast, private chat interface without requiring internet connectivity or external API services.

• 100% local inference — your data never leaves your machine
• Multi-architecture support: Llama, Qwen2, Qwen2.5, Qwen3, Qwen3 MoE, Mistral, Mixtral, DeepSeek, Yi, SmolLM2
• GGUF and SafeTensors model formats
• Hardware acceleration: CPU, CUDA (NVIDIA), Metal (Apple Silicon), Intel MKL, Apple Accelerate
• Streaming text generation
• Multi-language UI: English, Russian, Brazilian Portuguese
• Modern interface built with Svelte 5 and Tailwind CSS

• Node.js (for frontend build)
• Rust toolchain (for backend)
• For CUDA: NVIDIA GPU with CUDA toolkit
• For Metal: macOS with Apple Silicon

# Install dependencies
npm install

# Run with CPU backend
npm run tauri:dev:cpu

# Run with CUDA backend (NVIDIA GPU)
npm run tauri:dev:cuda

# Platform-aware development
npm run app:dev

# Build with CPU backend
npm run tauri:build:cpu

# Build with CUDA backend
npm run tauri:build:cuda

npm run lint          # ESLint
npm run lint:fix      # ESLint with auto-fix
npm run check         # Svelte type checking
npm run format        # Prettier formatting
npm run test          # Vitest tests

cargo clippy          # Linting
cargo test            # Unit tests
cargo audit           # Security audit

1. Build or download the application
2. Download a compatible GGUF or SafeTensors model (e.g., from Hugging Face)
3. Launch Oxide Lab
4. Load your model through the interface
5. Start chatting

• Windows, macOS, or Linux
• Minimum 8 GB RAM (16+ GB recommended for larger models)
• For GPU acceleration:
  • NVIDIA: CUDA-compatible GPU
  • Apple: M1/M2/M3 chip (Metal)
  • Intel: CPU with MKL support

Architectures with full support:

• Llama (1, 2, 3, 4), Mistral, Mixtral, DeepSeek, Yi, SmolLM2, CodeLlama
• Qwen2, Qwen2.5, Qwen2 MoE
• Qwen3, Qwen3 MoE

Formats:

• GGUF (quantized models)
• SafeTensors

• All processing happens locally on your device
• No telemetry or data collection
• No internet connection required for inference
• Content Security Policy (CSP) enforced

This project is built on top of excellent open-source work:

• Candle — ML framework for Rust (HuggingFace)
• Tauri — Desktop application framework
• Svelte — Frontend framework
• Tokenizers — Fast tokenization (HuggingFace)

See THIRD_PARTY_LICENSES.md for full dependency attribution.

Apache-2.0 — see LICENSE

Copyright (c) 2025 FerrisMind