TurboGPT is tiny byte-level GPT training written in CUDA C++ and released under MIT. You build it on Linux or NixOS with nix-build, or on Windows with Visual Studio 2022 C++ tools and CUDA 13.4, passing the compute capability for your GPU. Running it trains on a plain text file and stores a checkpoint with model, optimizer, scheduler and trainer state, resumable with --load, and writes TensorBoard-compatible logs with one report per batch. The published result is 2.5295 bits per byte on the hn1g set after 1.5 billion training tokens.
Be the first to rate TurboGPT. Your feedback helps other users decide.
Compile for your GPU's compute capability on Linux or Windows with CUDA 13.4.
Point it at a text file and let it checkpoint model, optimizer, scheduler and trainer state.
Continue from a checkpoint and read the TensorBoard-compatible reports per batch.
Similar tools in the same category — compare side-by-side
gPTY is a desktop workspace built on Rust and Godot. It gives you a tiling grid of panes — terminals...
Search2o is a stateless, asynchronous Python server that runs agents inside your organization, with ...
GitHub Copilot is a widely used AI coding assistant that suggests code directly inside VS Code, JetB...
Guides and comparisons about code tools
Ten launches that keep the model, the index and the facts on your own machine, from search that admits no match to GPU training in CUDA.
Discover the best AI code tools for DevOps engineers in 2026. Infrastructure as code, CI/CD optimization, monitoring and incident response compared.
AI coding assistants are transforming software development. Should developers be worried? An honest look at what AI changes and what it cannot replace.
Discover the best AI code tools for CI/CD pipelines in 2026. Compare automation, testing, deployment, and monitoring tools for faster delivery.
Discover the best AI code tools for API development in 2026. Compare tools for API design, documentation, testing, monitoring, and security automation.
