Build a private, unlimited ChatGPT alternative that runs entirely on your machine with zero cloud dependency.

Why Replace ChatGPT?

ChatGPT is impressive, but every message you send is stored on OpenAI's servers. Your conversations are used for training. You are subject to content filters, rate limits, and a $20/month subscription. For many use cases—journaling, brainstorming, code review, research—a local alternative gives you the same quality with none of the trade-offs.

What You Will Build

By the end of this guide you will have:

Ollama serving a language model (Llama 3.1, Qwen 3, or Mistral)
Open WebUI providing a polished chat interface with conversation history
Everything running on your machine with zero internet dependency

Install in One Click

Open Thumper-Run, go to the App Catalog, and search for Open WebUI. Click Install. Thumper will install both Open WebUI and Ollama as dependencies, configure the connection between them, and pull a default model. The install pipeline handles all of this automatically.

Choose Your Model

Ollama supports dozens of models. Here are the best options by VRAM:

8 GB VRAM or less: Llama 3.1 8B (Q4_K_M) – fast, capable, excellent at general chat
12 GB VRAM: Qwen 3 14B (Q4_K_M) – stronger reasoning, multilingual
16–24 GB VRAM: Llama 3.1 70B (Q4_K_M) – near-GPT-4 quality
CPU only: Llama 3.1 8B (Q4_K_M) at ~15 tokens/second on modern hardware

Thumper-Run's model management system shows which models fit your hardware before you download them.

Using Open WebUI

Open WebUI looks and feels like ChatGPT. You get:

Conversation threads with full history
System prompts to customize the AI's personality
File uploads for document Q&A (RAG)
Code highlighting with copy buttons
Model switching mid-conversation

With Ollama and Open WebUI configured for local-only operation and network integrations disabled, conversation data stays in the local Open WebUI database. Enabling connectors, remote models, sharing, or sync changes that boundary.

Performance Tips

Quantization matters. Q4_K_M is the sweet spot between quality and speed for most users.
Context length affects VRAM usage. Stick to 4096 tokens unless you need longer conversations.
Concurrent requests are possible but each one uses additional VRAM. Close unused conversations.

See the Ollama FAQ for more tuning advice.

Privacy Boundary

The local-only setup keeps inference and Open WebUI's own database on the machine. Ordinary local disk controls still apply. Current Thumper relay-assisted sync is server-assisted and is not a private E2EE guarantee. Review app network permissions and telemetry settings before use.

Ready to try it? Download Thumper-Run free →