[ 00 ] // open models · local runtime

Run open models
on your own GPU_

Chat, coding agents, speech and photo-to-3D on your own machine. No account, no monthly bill, and nothing you type ever leaves your computer.

Get ArcFlare /models for my gpu

Never run a model before? The ten-minute guide.

unit arc-01 // ascii online

move your cursor · click it · leave it alone for a bit

[ vs ]// the whole point

Their cloud, or your machine.

Every big-lab chatbot runs on someone else's computer. Here's what that costs you, and what you get back by running open models yourself.

✗ big-lab cloud ai

Rented, watched, metered.

✓ open models, on your machine

Yours, private, free.

your data

✗Every prompt, file and line of code is sent to their servers, under your account.

✓Nothing leaves your machine. There's no server to send it to, and no account.

training

✗Your conversations can be used to train their next model, unless you find the opt-out.

✓Your data trains nothing. It's a file on your own disk.

records

✗Chats are stored on their side, where reviewers and legal requests can reach them.

✓The only copy is yours. Delete it and it's gone.

cost

✗$20 to $200 a month, or a per-token bill that grows with every agent loop.

✓Free. Leave an agent running all night and you pay for the electricity.

limits

✗Rate limits and usage caps, usually right in the middle of something.

✓No caps, no queue, no “try again in 3 hours”.

offline

✗Down when they're down, and nothing at all without internet.

✓Works on a plane, in a cabin, during an outage.

control

✗They can swap, water down or retire the model whenever they like.

✓Open weights: the model you downloaded stays exactly as it is.

The open models are good now: Qwen, Llama, Gemma, DeepSeek, gpt-oss. They run on the card you already own.

Take it local

[ 01 ]// get it

Terminal or app. Same engine, same models.

[ cli ] // in a terminal

One line to install.

Needs Node.js 18+ and nothing else. It finds your llama.cpp and sizes every model to the VRAM you have.

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

01arcflare shoppick a model that fits your GPU
02arcflarechat with it
03arcflare agenta coding agent, all local

MetalCUDAROCmVulkandocs →

[ app ] // no terminal needed

Or get the app.

Chat, a studio for images, 3D and speech, and one-click coding agents. Windows, macOS and Linux.

Download for Windows· 10 / 11 · x64 installer

Free, MIT licensed, no account. All releases and notes · Source

[ 02 ]// what it does

Not just a chatbot.

agents

Chat and code

A coding agent with tools, MCP and skills, or OpenCode, Hermes and Codex pointed at your model.

❯ arcflare agent3d

Photo → 3D

One picture in, a .glb out for Blender or Godot. No upload, no credits.

❯ arcflare gen 3d chair.pngspeech

Speech

Natural voices, tone control and cloning from a short clip, in ten languages.

❯ arcflare gen tts "hi"remote

From your phone

Scan a QR code and use your machine from anywhere. The model stays home.

❯ /rc

[ 03 ]// rangebest pick for my gpu →

What runs on what.

16 GB

RTX 4060 Ti 16G · 4080 · 5070 Ti

gemma39 GB mistral-nemo-12b9 GB hunyuan3d-2.110 GB phi410 GB qwen3-14b10 GB deepseek-r1-distill-qwen-14b10 GB qwen2.5-coder-14b-instruct10 GB phi-4-reasoning-plus11 GB starcoder211 GB deepseek-coder-v2-lite12 GB gpt-oss14 GB codestral15 GB ernie-4-5-21b-a3b15 GB devstral2-small16 GB mistral-small3.116 GB devstral16 GB magistral-small-250916 GB

24 GB

RTX 3090 · 4090 · 7900 XTX

mistral-small-3.2-24b17 GB qwen3.518 GB qwen3-coder19 GB gemma3-27b19 GB glm4.7-flash20 GB qwen3-30b-a3b-instruct-250720 GB qwen3-30b-a3b-thinking-250720 GB qwen2.5-coder-32b21 GB qwen3-32b21 GB deepseek-r1-distill-qwen-32b21 GB granite-4.0-h-small21 GB olympiccoder22 GB nextcoder22 GB qwen2.5-vl-32b-instruct22 GB seed-oss-36b-instruct23 GB

64 GB

workstations · 64 GB Macs

mixtral28 GB kimi-dev42 GB llama3.3-70b44 GB qwen3-coder-next52 GB glm4.5-air64 GB llama4-scout64 GB gpt-oss-120b64 GB

64+ GB

128 GB Macs · servers

glm4.6220 GB kimi-k2250 GB

[ 04 ]// faq

Questions, answered.

Something missing? Read the beginner's guide or ask us.

What is ArcFlare?

ArcFlare is a free, open-source tool for running AI models on your own computer. It comes as a command-line tool and a desktop app for Windows, macOS and Linux, and covers chat, coding agents, image-to-3D and text-to-speech, using llama.cpp under the hood.

Is ArcFlare free?

Yes. ArcFlare is MIT licensed, has no account, no subscription and no per-token cost. The models it runs are open-weight models with their own licenses, listed on each model page.

Does anything I type leave my computer?

No. Models run on your own GPU or CPU, and ArcFlare has no telemetry. It only goes online to download models you choose, check for updates, and relay remote control when you turn it on.

What hardware do I need?

Any recent PC or Mac works. Small models run on a CPU with 8 GB of RAM; a GPU with 8–16 GB of VRAM runs strong chat and coding models; Apple silicon Macs use their unified memory. ArcFlare checks what fits before it downloads or loads anything.

Do I have to use the terminal?

No. ArcFlare Desktop does everything in a window: pick a model, chat, generate images, 3D and speech, and open coding agents. The CLI is there if you prefer it, and both share the same models.

Which coding agents does it work with?

OpenCode, Hermes, Codex and ArcFlare's own agent, all pointed at your local model with one command. ArcFlare can also update them with arcflare harness update.

Can I use it from my phone?

Yes. Type /rc in a session and open the link on your phone, from any network. You can read along, send messages and approve tool calls, while the model and your files stay on your machine.

Run open modelson your own GPU_