How to Set Up a Free, Uncensored AI Companion Locally (No GPU)

You can test the model right now with ollama commands. But the terminal forgets every message, and what you installed is a blank model with no name and no personality.

So I built a small chat page for you. It runs in your browser, talks straight to your local Ollama, and keeps everything local. You set a name, an avatar, and a personality once.

Everything above assumes CPU only, which is deliberately the lowest common denominator: it runs on a decade-old laptop. If you have a GPU sitting idle, a bigger local model becomes realistic, and the appeal is the same as the CPU setup, nothing leaves your machine, plus noticeably better memory and coherence in a long roleplay session than a 4B model can manage.

The model worth pointing at here is Qwen 3.8 27B, one tier up from the abliterated 4B this guide installs by default. At Q4_K_M quantization (about 14GB) it fits a single 16GB-VRAM consumer GPU alongside 32GB of system RAM, running around 25 to 35 tokens per second. Push to Q3_K_M (about 12.5 to 13GB) on that same 16GB card and throughput rises to roughly 41 tokens per second, with a small quality cost. Running it on system RAM instead of a GPU (the BF16 full-precision weights, about 55GB, needing 64GB+ of RAM) drops throughput to around 4 tokens per second, a bandwidth bottleneck that makes it slower than the CPU-only 4B setup above rather than an upgrade from it, so this model is worth running only with a GPU to offload onto.

On cost, running Q3/Q4 on a single RTX 3090/4090-class card at roughly 250W draws about 0.37 US dollars per million output tokens at typical US residential electricity rates, against roughly 3.00 US dollars per million output tokens on a comparable FP8 cloud endpoint at lower throughput. We confirmed this ourselves in a roleplay session on the abliterated build, not just the coding/agentic benchmark it was first measured on: same speed per tier, same cost, nothing about running it as a companion instead of a coding assistant changes the numbers.

Local is the right default for privacy, and it is the only setup where you can keep chatting offline, on CPU or GPU. But it is not automatically the better experience. Even a GPU-accelerated local model tops out below what you can rent by the token, and in long conversations that gap shows up as thinner memory and flatter replies.

Pick local when privacy is the point, when you want zero running cost, or when you want it to work offline. Pick a hosted API when you want a larger model than any of your own hardware can hold.

Which hosted models allow uncensored roleplay is its own question, so we wrote up which hosted APIs actually allow uncensored roleplay to give some solutions.