tabby

Start Tabby server with GPU acceleration

$ tabby serve --model [StarCoder-1B] --device cuda

$ tabby serve --model [StarCoder-1B] --chat-model [Qwen2-1.5B-Instruct] --device cuda

$ tabby serve --model [StarCoder-1B] --device cpu

$ docker run -it --gpus all -p 8080:8080 -v $HOME/.tabby:/data tabbyml/tabby serve --model [StarCoder-1B] --device cuda

$ tabby serve --model [StarCoder-1B] --port [8080]

tabby serve [--model name] [--chat-model name] [--device type] [--port port]

tabby is a self-hosted AI coding assistant that provides code completion, inline edits, and chat capabilities. Unlike cloud-hosted alternatives, Tabby runs entirely on your own infrastructure, giving you full control over models, data, and costs.
The serve command starts the Tabby API server, which exposes an OpenAPI-compatible interface for IDE extensions and other clients. The server supports multiple code completion models including StarCoder, CodeLlama, and CodeGen families.
Tabby is optimized for consumer-grade GPUs and supports NVIDIA CUDA for Linux/Windows and Apple Metal for macOS. CPU-only mode is available for environments without GPU acceleration, though with reduced performance.
Data is stored in ~/.tabby by default, including model weights, configuration, and indexed code repositories. The server provides a web UI at the configured port for administration, model management, and repository indexing.

--model name

Code completion model to use (e.g., StarCoder-1B, CodeLlama-7B).

Conversational AI model for chat features (e.g., Qwen2-1.5B-Instruct).

Hardware acceleration: cuda (NVIDIA GPU), metal (Apple M1/M2), cpu.

Port to expose the API server. Default: 8080.

Display help information.

GPU acceleration requires appropriate drivers (CUDA toolkit for NVIDIA, Metal for Apple Silicon). Model downloads can be several gigabytes depending on the selected model. Self-hosting requires adequate hardware resources; recommended minimum is 8GB VRAM for GPU mode or 16GB RAM for CPU mode.

Tabby was created by TabbyML and released as open-source in 2023, positioning itself as a self-hosted alternative to GitHub Copilot. The project gained traction among enterprises and developers seeking data privacy and cost control. Version 0.24 in February 2025 added LDAP authentication, and version 0.30 in July 2025 introduced GitLab merge request context indexing.

ollama(1), llamacpp(1), copilot(1)