Run any local LLM engine,
auto-tuned to your GPU

Bring your own llama.cpp fork. No compiling. No Electron. No Python. Point Claude Code at your own machine in one command — fully offline.

$ npx turbollm

Don't have Node? One command installs it.

Windows: winget install OpenJS.NodeJS.LTS · macOS: brew install node · Linux: your package manager or nodejs.org (v22+). Then open a new terminal and re-run npx turbollm.

Runs llama.cpp, ik_llama.cpp, TurboQuant, KoboldCpp, llamafile, MLX, vLLM, SGLang — or any community fork.

Get Started What can I run? Browse models

localhost:6996

TurboLLM chat with a local model, showing real generation speed of 85.9 tokens per second, dark mode

TurboLLM chat with a local model, showing real generation speed of 85.9 tokens per second, light mode

2.2×

Faster than other inferences

~7MB

npm package size

Engine types supported

Telemetry collected

Video: Run Bonsai 27B with 1 click on your GPU, no manual build needed

Run Bonsai 27B with one click — no manual build

Video: Run vLLM on Windows via WSL2 — real setup with TurboLLM

Run vLLM on Windows via WSL2 — the real setup More on the @TurboLLM channel →

Any engine, including forks

Point it at any llama.cpp-compatible binary — a build you compiled, a community fork, or the one it auto-provisions for your GPU. The fastest community innovations land in forks first.

Auto-tuned to your hardware

Benchmarks on load, derives fast defaults, and shows a VRAM-fit verdict before you load — no more flag guessing.

Real tokens/sec, never faked

Speed in the model list is measured on your machine from actual generation — live while you chat, and remembered per model.

Drop-in APIs

OpenAI and Anthropic-compatible — so Claude Code and every existing tool work unchanged.

Offline-first & private

No account, no backend, no internet, no telemetry. Your prompts, chats, files, and keys never leave your machine.

Use from any device

The UI runs in the browser, so any phone, tablet, or laptop on your LAN can use the model on your GPU box.

Local coding agent

Read, edit, run commands, and report diffs in a real repo folder — built on pi's SDK, running on your own GPU. No cloud key, no Electron.

localhost:6996

TurboLLM Engines screen showing hardware-fit marks and grounded pros/cons for llama.cpp, ik_llama.cpp, llamafile, KoboldCpp, TurboQuant, and vLLM, dark mode

TurboLLM Engines screen showing hardware-fit marks and grounded pros/cons for llama.cpp, ik_llama.cpp, llamafile, KoboldCpp, TurboQuant, and vLLM, light mode

See what actually fits your GPU

Every engine card carries a hardware-fit verdict for your exact GPU, grounded pros/cons instead of marketing copy, and whether it's a one-click install or a build from source. Bring any llama.cpp fork — the same catalog covers vLLM, MLX, and TurboQuant too.

See in docs

localhost:6996

TurboLLM Models screen listing GGUFs with measured tokens/sec and VRAM-fit verdicts, dark mode

TurboLLM Models screen listing GGUFs with measured tokens/sec and VRAM-fit verdicts, light mode

Models, without the busywork

Point it at the GGUF folders you already have — LM Studio folders or manual downloads — and it reuses them in place, no re-downloading. Every model shows a VRAM-fit verdict and its measured tokens/sec, from actual generation on your GPU, so you pick a quant that fits before you commit. One click to Load.

Browse the models hub

localhost:6996

TurboLLM Customize screen with editable built-in agents, a custom system prompt, and a skill/tool allow-list, dark mode

TurboLLM Customize screen with editable built-in agents, a custom system prompt, and a skill/tool allow-list, light mode

Agents you actually control

Edit any built-in agent's system prompt, skills, or tool access in place — Reset undoes it — or build your own from a name, a prompt, and a checklist. A shared Skills library (Claude-style SKILL.md) any conversation can turn on, plus an MCP servers list your agents can draw tools from.

See in docs

localhost:6996

TurboLLM Code screen showing a local coding agent session with tool-call lines, real stats footer, and a git diff, dark mode

TurboLLM Code screen showing a local coding agent session, light mode

A local coding agent that works in your repo

Open Workspace → Code and the agent starts working in a real folder — reading files, editing code, running commands, and reporting diffs. Built on pi's SDK with a dense terminal feel, real stats footer, and full tool access including MCP tools and skills. Optional worktree isolation keeps changes clean.

See in docs

localhost:6996

TurboLLM Developer screen showing one-command CLI setup cards and API key management, dark mode

TurboLLM Developer screen showing one-command CLI setup cards and API key management, light mode

Built for the tools you already use

OpenAI- and Anthropic-compatible endpoints on the same port, API keys for LAN sharing, and a one-command CLI hookup — turbollm launch claude points Claude Code at your own GPU with no cloud key.

See all integrations

Qwen3.6-35B-A3B · 200K	TurboLLM	LM Studio	Speed-up
official llama.cpp — q4_0	74.7 t/s	61.0 t/s	1.2×
official llama.cpp — q8_0	72.3 t/s	~66 t/s	1.1×
TurboQuant fork — turbo4	24.6 t/s	11.4 t/s	2.2×

	TurboLLM	LM Studio	Ollama	Open WebUI
Run any engine / forks	✓	✗	✗	✗
Benchmark-based auto-tune	✓	◐	◐	✗
Measured t/s in model list	✓	◐	◐	✗
Anthropic API → Claude Code	✓	✓	✓	✗
OpenAI-compatible API	✓	✓	✓	◐
Lightweight (no Electron / Python)	✓	✗	✓	✗
Offline-first, no telemetry	✓	◐	✓	✓

Want the detail? TurboLLM vs LM Studio · TurboLLM vs Ollama — honest, table-driven, updated July 2026.

Get started in one command

No installation, no setup. Just run it.

npx turbollm

Or install globally: npm install -g turbollm · New here? Read the getting-started guide →

Run any local LLM engine,
auto-tuned to your GPU

200k context. 27B model. 16 GB of VRAM.