# TurboLLM

A local-LLM runner that auto-tunes launch flags to the user's GPU, serves models behind a web UI and an OpenAI or Anthropic-compatible API, and can run any llama.cpp-compatible engine, including community forks, instead of one fixed build.

TurboLLM starts a local daemon with a single command, opens a browser-based UI, and serves installed models over an API any compatible tool can call, including pointing Claude Code at the user's own machine. Rather than shipping one blessed inference runtime the way some local-LLM tools do, it can point at any llama-server-compatible binary, whether a build the user compiled themselves, a community fork, or one it auto-provisions for the detected GPU, probing the binary's actual capabilities and adapting the interface to them.

It benchmarks the machine on load, derives launch flags such as context size, GPU layer count and KV cache type automatically, and reports a VRAM-fit estimate before a model is loaded rather than requiring the user to guess. It is distributed under a source-available license that is free for personal and internal business use and converts each release to Apache-2.0 two years after it ships, and it is installed and started with a single npx command with no Electron, Python or manual compilation step required.

Website: https://turbollm.dev