[ docs ] // Documentation

ArcFlare

Run open models on your own GPU and point the tools you already use at them. Chat and code with GGUF language models, turn pictures into 3D meshes, and reach the whole thing from your phone. No cloud, no account, and no npm dependencies.

Install#

The installer needs no root or admin. If you don't have Node.js 18 or newer, it installs a private copy (checked against nodejs.org's checksums) just for ArcFlare. It then downloads llama.cpp, the engine that runs models, for your machine and checks it against the SHA-256 GitHub publishes for it; an engine you already have is left alone. Skip that with --no-engine ($env:ARCFLARE_NO_ENGINE=1 on Windows).

macOS / Linux

curl -fsSL https://arcflare.net/install.sh | sh

Windows (PowerShell)

irm https://arcflare.net/install.ps1 | iex

Also download a model that fits your GPU (opt-in; it is a large download)

curl -fsSL https://arcflare.net/install.sh | sh -s -- --model

$env:ARCFLARE_MODEL=1; irm https://arcflare.net/install.ps1 | iex

It picks the largest general or coding model that fits your free GPU memory and is fine for commercial use; the same choice is arcflare shop best (add --dry-run to see it first).

The desktop app instead (Linux: pacman, apt or AppImage; macOS)

curl -fsSL https://arcflare.net/install-desktop.sh | sh

The desktop app on Windows (checks the download, skips the SmartScreen warning)

irm https://arcflare.net/install-desktop.ps1 | iex

From source

git clone https://github.com/Hakeperty/ArcFlare-Code.git && cd ArcFlare-Code && npm link

Then check the engine, your GPU and every harness at once:

arcflare doctor

  ✓ llama-server  C:\Users\you\llamacpp\vulkan\llama-server.exe
  ✓ backend       vulkan  AMD Radeon(TM) 8060S Graphics (active)
  ✓ memory        balanced profile
  ✓ device memory 45.4 GB free
  ✓ models        4 found
  ✓ OpenCode       ...\node\opencode
  ✓ mcp server    29 tools · arcflare mcp

No llama.cpp yet?

The installer fetches it for you. If that step was skipped or failed (offline, a GitHub rate limit), one command downloads the right prebuilt build for your machine (Metal on macOS, Vulkan or CPU on Linux, CUDA or Vulkan on Windows) and sets it up: arcflare engine install. Already have one somewhere else? Point ArcFlare at it with arcflare set-engine /path/to/llama-server.

Your first run#

Start with no arguments. The menu walks you through a harness, a model and a context size, and remembers your choices for next time:

arcflare

No models yet? Open the shop: it shows which models fit your GPU and downloads the one you pick.

arcflare shop

Or pull a GGUF from Hugging Face directly, then chat with it:

arcflare pull unsloth/Qwen3.6-35B-A3B-GGUF:Q5_K_XL

arcflare run qwen3.6-35b-a3b

Inside a session, /rc connects your phone, /update installs a new version, and /bye leaves.

Prefer an app?#

ArcFlare Desktop does all of this in a window for Windows, macOS and Linux: chat, a studio for images, 3D and speech, flows, one-click harnesses, and a QR code to use it from your phone. It shares models and settings with the CLI, and updates itself.

What changed lately is on the changelog, straight from the repositories.

Brand new to this?#

Read New to ArcFlare? Here's everything you need to know. It covers what your computer needs, picking a model, the commands you will use, and the jargon, in plain words.

Where to next#

01Running modelsModels & context →Where models are found, pulling them, and how the context window is sized.02Running modelsGPU & performance →Backends, memory profiles, batch size, and measuring what runs best.03Running modelsHarnesses →OpenCode, Hermes, Codex and the built-ins, pointed at your model.04Agents & toolsThe agent →A coding agent built around a local model's prompt cache: tools, MCP, skills.05Agents & toolsMachine server →An MCP server that runs, opens, sees, clicks and tests on your computer.06Beyond chat3D generation →Hunyuan3D and TripoSR: a photo in, a .glb mesh out, on your GPU.07Beyond chatSpeech →Text-to-speech with Qwen3-TTS, Kitten TTS 2, Kokoro, Chatterbox and more: voices, tone, cloning, saved voices.08Beyond chatRemote control →Type /rc and use the session from your phone, anywhere.09ReferenceCommands →Every command, slash command, environment variable and file.10ReferenceUpdates →Update notices, /update, and updating machines with no internet.11ReferenceTroubleshooting →The problems people actually hit, and what fixes them.