Neura Market
TurboPilot logo

TurboPilot

Free

A self-hosted copilot clone which uses the library behind llama.cpp to run the 6 billion parameter Salesforce Codegen model in 4 GB of RAM.

FreeFree tier
Inputs: codeOutputs: code
Type
Open Source
Visit Website GitHub (40 stars)

About TurboPilot

TurboPilot is an open-source, self-hosted AI code completion engine that runs locally on CPU, leveraging the llama.cpp library to run quantized large language models. It supports multiple models including Salesforce Codegen (6B parameters), StableCode 3B, WizardCoder, StarCoder, and SantaCoder, enabling code completions in various programming languages. Designed as a local alternative to GitHub Copilot, TurboPilot can run on machines with as little as 4GB of RAM by using quantized GGML models. The server listens on port 18080 and integrates with editors via the fauxpilot extension. Note: The project was deprecated and archived as of September 2023, with slower autocompletion compared to commercial solutions.

Key Features

Self-hosted code completion engine running entirely on local CPU
Uses llama.cpp for efficient inference with quantized models
Supports multiple state-of-the-art models: Salesforce Codegen, StableCode 3B, WizardCoder, StarCoder, SantaCoder
Runs with as little as 4GB RAM using quantized GGML models (e.g., StableCode 3B)
Provides a server on port 18080 compatible with fauxpilot VS Code extension
Open source (MIT license) with option to download pre-converted models from Hugging Face

Pros & Cons

Pros
  • Fully self-hosted, ensuring data privacy and no external API calls
  • Low hardware requirements (4GB RAM) for running code completion models
  • Supports multiple modern code completion models with 'fill in the middle' support
  • Open source and free to use, modify, and distribute
  • Easy setup via pre-converted models and Docker
Cons
  • Project is deprecated/archived as of September 2023 with no further development
  • Autocompletion speed is slow compared to commercial solutions like GitHub Copilot
  • Documented as a proof of concept, not a stable production tool
  • Requires manual setup and configuration of models and server

Best For

Offline AI-powered code completion for developersSelf-hosted alternative to GitHub Copilot for privacy-conscious teamsCode assistance in restricted or air-gapped environmentsExperimenting with local large language models for code generation

FAQ

What hardware do I need to run TurboPilot?
TurboPilot can run on machines with 4-8GB RAM using quantized models like StableCode 3B. For higher performance, 16+GB RAM or a discrete GPU / Apple Silicon is recommended for models like WizardCoder.
How do I get the models?
Pre-quantized models can be downloaded from Hugging Face. Alternatively, you can convert models yourself using provided scripts. See MODELS.md for a full catalogue.
How do I start the TurboPilot server?
Download the binary or build from source, then run: ./turbopilot -m starcoder -f ./models/santacoder-q4_0.bin. The server starts on port 18080 by default.
Is TurboPilot still actively maintained?
No, the project was deprecated and archived on September 30, 2023. The creator recommends using more mature alternatives.
Which code editors are supported?
TurboPilot uses the fauxpilot extension protocol, so it works with editors that support that extension, such as VS Code.