Open source alternatives to ChatGPT
A self-hosted replacement for ChatGPT is a three-part stack: a chat interface such as Open WebUI, a model runner such as Ollama, and an open-weight model such as Qwen or Llama. On a 12 GB GPU or a recent Mac, that covers most of what a $20/month Plus subscription does, and every conversation stays on your hardware.
Updated July 11, 2026
The short answer
| You want | Run this |
|---|---|
| chatgpt.com, but on your own machine |  Open WebUI+ Ollama |
| One desktop app, no Docker, no server |  Jan |
| Chat over your own documents |  AnythingLLM |
| One server for a whole team |  vLLM+ LibreChat |
| An assistant that knows your notes |  Khoj |
Why people replace ChatGPT
OpenAI's consumer pricing as of July 2026 runs from a free tier limited to roughly 10 messages every 5 hours, through Go at $8/month (ad-supported in the US), Plus at $20/month, and Pro tiers at $100–200/month. Plus buys around 160 messages every 3 hours on the main models, which is plenty for chat and tight for heavy coding sessions.
For most people who switch, the sticking point is data rather than price. Consumer plans use your conversations for model training unless you find the opt-out under Settings → Data Controls. And 2025 showed that retention policy can change without your consent: a preservation order in the New York Times lawsuit forced OpenAI to keep all consumer chats, including deleted ones, from May 2025 until the obligation was lifted that September. Deleted chats are back to a 30-day purge today. The lesson people took from the episode still holds: chats on someone else's server are governed by someone else's legal obligations.
The stack, layer by layer
Nobody ships a single "open source ChatGPT". The replacement is three layers that talk to each other over one shared API convention, which means you can swap any layer later without redoing the others.
The interface
The part you look at. All three speak the OpenAI-compatible API, so they work with any runner below.
The default. It reproduces ChatGPT's layout closely enough that nobody at home will ask questions, and it grew RAG, web search, and multi-user support long ago. If you want chatgpt.com on your own domain, start here.
Closest to feature parity, with agents, a code interpreter, and connections to every major provider. That last part makes it a natural hybrid setup: local models for private work, an API key for the hardest questions.
The best-looking of the three, with a plugin marketplace. Pick it on design taste.
The runner
The part that executes the model on your hardware.
One-line install, a model registry that works like Docker's, and an OpenAI-compatible API. The runner most people should start with, even though it carries overhead compared to raw llama.cpp.
The C/C++ engine underneath half the local-AI ecosystem, Ollama included. Run it directly when you want control over quantization and context size that Ollama's defaults hide from you.
Built for throughput. When several people share one server, its continuous batching serves them concurrently instead of queueing them.
The model
Open-weight model families you can download and run. Quality per gigabyte improves every few months, so check release notes before committing VRAM.
The community default in 2026, especially for coding. The 8B and 14B variants are what most 8–12 GB GPUs end up running day to day.
Meta's model line, still the most polished general-purpose chat behavior in the open-weight world.
OpenAI's own open-weight models. The option for people who trust the ChatGPT name but need local deployment.
Shortcuts and specialists
Projects that collapse the stack into one install, or go past plain chat.
A desktop app that bundles interface and runner with no Docker involved. The shortest path from download to a first local conversation.
A workspace for chatting over your own documents, with RAG built in rather than bolted on.
A personal assistant over your notes and files that can also run scheduled automations.
What your hardware can run
The numbers below assume Q4_K_M quantization, the balance point most of the community settles on. Quantizing to Q4 costs a little quality and roughly halves memory needs; a 70B model still wants 35–45 GB even then. Longer context windows add memory on top, so treat these as floors.
| Hardware | What fits | How it compares |
|---|---|---|
| 8–12 GB VRAM (RTX 4060-class) | 8–14B models at Q4, 40–80 tokens/s | Around ChatGPT's free tier, often better at coding |
| 16–24 GB VRAM (RTX 3090 / 4090) | 27–35B models, including efficient MoE variants | Matches Plus on coding and math, thinner on long reasoning |
| Apple Silicon, 32 GB+ unified memory | 30B-class models run comfortably | Same ballpark as the 24 GB GPU row |
| 48 GB+ (multi-GPU or Mac Ultra) | 70B dense and large MoE models | Competitive with paid tiers on most work |
What you give up
Maintenance is real. CUDA and ROCm driver updates break things, runners ship with default context windows of 2–4k tokens when useful work starts at 8k, and there is always one more quantization variant to try. A Hacker News commenter described running a local model as "adopting a pet dragon. Fun at first, but then it keeps eating all my GPU and still refuses to clean up its own context window." People who enjoy that kind of tinkering stay; people who wanted an appliance drift back within a month.
Quality gaps show up in specific places. Local models handle everyday chat and coding well at 27B and up, and then stumble on careful document parsing and long multi-step reasoning, which is exactly why so many switchers end up hybrid, with local models for private work and a hosted model for the hard problems. There is also no integrated equivalent of ChatGPT's voice mode, and image generation means standing up a second stack from the generative media section of the registry.
Then the money. A new RTX 4090 build lands near $2,900, which buys twelve years of Plus, and a GPU that idles 22 hours a day is a bad investment compared to paying per token. The economics work when you already own the hardware, when professional use keeps hitting rate limits, or when the data genuinely cannot leave your network. Privacy is the honest reason to self-host; saving money usually is not.
Questions people actually ask
Can an open model really match ChatGPT?
Depends on the size you can run. An 8–14B quantized model lands around the free tier's quality and often beats it at coding. A 27B+ model on a 24 GB card matches Plus on coding and math in both benchmarks and user reports, while long multi-step reasoning and careful document analysis still favor the hosted frontier models. The gap has narrowed every year since 2023.
What is the fastest way to try this?
Install Jan if you want a desktop app, or install Ollama and type "ollama run qwen3:8b" in a terminal. Either way you will know within an hour whether local quality clears your bar, before touching Docker or buying hardware.
Do I need a GPU?
For daily use, effectively yes. CPU-only inference works but is slow enough that you will stop using it. Apple Silicon is the exception: unified memory runs 30B-class models well on M-series Macs with 32 GB or more.
Does self-hosting actually fix the privacy problem?
Mechanically, yes. The model runs on your hardware and conversations never leave it, so there is no retention policy to read and no training default to opt out of. On consumer ChatGPT plans, conversations are used for model training unless you disable it under Settings → Data Controls.
Every project above is one row in the Awesome Open Source AI registry, which resyncs twice a day from the curated GitHub list. The User Interfaces & Self-hosted Platforms category holds a dozen more chat UIs and platforms that didn't make this shortlist.
