I use ChatGPT and Claude every day, but watching a 27B model download onto my own computer felt different. I opened Terminal, launched the model, and watched it answer from my machine without a cloud chat or per-message meter.
Then I found Qwen3.8-27B-OBLITERATED, a third-party modification whose author says the model’s refusal behavior and safety lectures were removed. That is where local AI starts to feel like the Wild West.
The technology impressed me, and the hardware demands surprised me. The modified model can produce material that a mainstream assistant would refuse, including content that is false, offensive, dangerous, or unlawful.
My recommendation for a beginner is easy: start with the official Qwen model. Learn how local AI works before you experiment with a guardrail-removed version.
The short version: Qwen 3.8 gives you privacy, control, and a powerful local workspace. You trade cloud convenience for hardware costs, slower performance, and more responsibility.
The good
Running a model locally changes what you can do with AI:
Keep many prompts and documents on your computer, as long as you do not connect the runner to outside services.
Reduce or replace some ChatGPT or Claude use for everyday drafting, summaries, brainstorming, and coding help.
Work after the model has downloaded, even without an internet connection.
Chat without a subscription message cap or per-message API charge.
Test compatible coding tools and supervised agents with less API usage.
Analyze private notes, PDFs, and business documents without uploading them to a cloud chat.
Choose the model, system instructions, context size, and interface that fit your workflow.
Build repeatable local automations for routine work.
Keep a useful AI tool available if a cloud service has an outage or changes its pricing.
Learn how models, quantization, context, memory, and inference work by seeing the pieces yourself.
Local AI will not replace every cloud product. ChatGPT and Claude still offer polished interfaces, hosted tools, and large systems that a home computer may not match. A local model can cover a surprising amount of daily work once you accept those limits.
The bad
Qwen 3.8 is a large model. The Q4 build used in this guide is around 16.8 GB, and the download size tells only part of the story. The runner also needs memory for the model, your conversation, and the rest of your computer.
These are my conservative starting recommendations, not manufacturer guarantees:
Computer | Recommended starting point |
|---|---|
Mac | Apple Silicon, current macOS, 32 GB unified memory, and at least 25 GB free space |
Windows PC | Modern x64 processor, 32 GB system RAM, 16 GB or more GPU memory preferred, and at least 25 GB free space |
A 64 GB Mac or a Windows machine with 24 GB of GPU memory gives you more room. A 16 GB computer should start with a smaller model. CPU-only loading may work, but responses can feel slow.
Local AI also costs you time. You manage downloads, updates, storage, context limits, and troubleshooting. The model may lag behind the newest hosted systems. A cloud subscription can still make sense when you value speed and convenience more than control.
The ugly
“Uncensored” sounds exciting until the model gives a confident answer with no judgment behind it.
The OBLITERATED build is not an official Qwen release. Its author describes it as a research model with refusal behavior removed. Fewer refusals do not make the model more accurate. It can answer a bad request, invent facts, produce biased material, or generate instructions that create real risk.
Keep a guardrail-removed model inside a supervised local chat. Do not connect it to:
Your shell, files, passwords, accounts, or payment systems
An autonomous agent or a public internet endpoint
Do not use it with minors, confidential client records, medical information, or anything you cannot afford to expose or corrupt. Review each output before you use it.
The model author, Qwen, Hugging Face, Ollama, and LM Studio control their own software and repositories. I do not create, host, inspect, or guarantee their model files.
Install the official Qwen 3.8 model
This beginner path uses Ollama and a Q4_K_M GGUF conversion of the official Qwen base model. Q4 keeps the download and memory demand below the full 55 GB model while preserving useful quality.
1. Check your computer
Use the recommendations above before downloading anything. Keep other large apps closed during the first launch.
2. Install Ollama
Download Ollama from ollama.com/download. Choose macOS or Windows, finish the installer, and open the Ollama app once.
3. Open the command window
Mac: Press Command + Space, type Terminal, and press Return.
Windows: Press the Windows key, type PowerShell, and open Windows PowerShell.
4. Download and launch the model
Paste this command and press Return:
ollama run hf.co/lmstudio-community/Qwen3.8-27B-GGUF:Q4_K_MThe first run downloads the model. Keep the computer awake and leave enough storage space. Ollama opens a local prompt when the download finishes.
5. Run a harmless test
Paste this prompt:
Explain in three sentences what it means for an AI model to run locally.Your setup works if the model responds.
LM Studio gives you a visual interface:
Install LM Studio from its official download page.
Search for
Qwen3.8-27B.Choose a
Q4_K_MGGUF build from a repository you trust.Download it, load it into Chat, and begin with a modest context size.
Longer context uses more memory. Raise it after you know the model runs well on your computer.
If you still want the modified version
Read the complete OBLITERATUS model card before downloading it. The page identifies the model as a third-party modification, documents the available file sizes, and provides its current installation commands.
I am keeping the official model as the beginner path. Adults who choose the modified model should keep it local, supervised, disconnected from tools, and away from sensitive information.
Common problems
“ollama: command not found” Open the Ollama app, close Terminal or PowerShell, open a fresh window, and try again.
The model will not load Close memory-heavy apps and restart. Use a smaller quantization or a smaller model if it still fails.
The model runs too slowly Shorten the context and close other apps. CPU-only and partial GPU loading can take much longer than full GPU or Apple Silicon acceleration.
I want to remove the model Run ollama list, copy the exact model name, then run ollama rm <model-name>.
My take
Local AI is worth learning now. You gain privacy, control, and a way to build without paying for every interaction. You also become the system administrator and the safety layer.
Install the official model first. Give it a few harmless tasks. Compare the speed and quality with the cloud tools you already use. That will tell you where local AI belongs in your workflow.
The uncensored version shows how fast this field is moving and how little structure surrounds some of the most powerful releases. Treat it as an adult research tool, not a shortcut to better answers.
Want to learn and build alongside other AI builders? Join us here.
Sources
Disclaimer
This post is general education, not legal, cybersecurity, medical, financial, or professional advice. Marcin AI, AI Builders, and MT Solutions, LLC are not affiliated with Qwen, Alibaba, Hugging Face, Ollama, LM Studio, OBLITERATUS, or their contributors. We do not create, host, inspect, guarantee, or control third-party software or model weights. Models may produce inaccurate, offensive, unsafe, or unlawful content. You are responsible for downloads, system security, supervision, outputs, and compliance with applicable laws, licenses, workplace rules, and third-party terms.