Qwen3.8-27B OBLITERATED: Why the Same Model Gives Different Answers
Qwen is a family of AI models that people can download and run on their own computers. Qwen3.8-27B is one model in that family. It can take a written request and produce text, code, or an answer without sending the work to Qwen's own service, provided the computer and app can run it.
Qwen3.8-27B-OBLITERATED is not an official Qwen release. It is a copy changed by an independent developer who wanted the model to refuse fewer requests. The dramatic name “OBLITERATED” is the developer's label for trying to remove that refusal behaviour. It does not mean the model is more accurate, more capable, or certain to answer everything.
Update, 23 August 2026. The developer announced that the earlier downloads in the GGUF format — a common file format for local chat apps — had not been converted correctly, and asked anyone who downloaded one to replace it with the latest release, labelled V3. The developer says the full-size original files, published in a format called safetensors, were never affected. The download page no longer offers the versions prepared for Apple computers; the developer's notes say those are waiting on updates to another tool. The comparison described below used a saved copy of one of those earlier Apple versions, so it reflects that older download, not the current V3 files.
The same download gave two answers. The downloaded files, typed request, and answer settings stayed the same; only the extra directions added by the app changed.
Data behind this diagram
| Part | What it was | Why it mattered |
|---|---|---|
| Official base | Qwen/Qwen3.8-27B | The original model |
| Unofficial changed copy | OBLITERATUS version designed to refuse less | Not an official Qwen release |
| Downloaded files | Saved copy of the developer's second version for Apple computers, since removed from the download page | Exactly the same files were used both times |
| First setup | Sent only the request typed by the user | The model answered the request |
| Second setup | Included extra directions added by the app | The same request was refused |
What the unofficial version changed
The OBLITERATUS project looks for patterns inside an AI model that are associated with refusing a request, then changes the model to weaken those patterns. The developer says the goal is to reduce refusals while keeping the model's other abilities. Those performance claims come from the developer's own tests and should not be read as an independent guarantee.
The download page identifies the official Qwen3.8-27B model as its starting point. When the comparison was run, it offered several versions of the changed model for different programs and computers. The local comparison used a saved copy of the developer's second version prepared for Apple computers — a version the page has since removed. Both sides of the comparison used exactly the same downloaded files.
Why the same download gave two answers
What you type into a chat box is not always everything the model reads. Before an app asks the model for an answer, it may add its own hidden directions. These can describe the assistant's role, tell it when to refuse, or explain how it should use features built into the app. Another app may add different directions or none at all.
In this comparison, one setup sent the user's request without those extra app directions. The other included them. On one request for content the model would normally reject, the first setup answered and the second refused. The model files did not change. The visible request did not change. The difference was the extra text sent by the software around the model.
This is why two local chat apps can disagree even when both display the same model name. Each app may format the conversation differently, add its own directions, or turn an optional answer mode on or off. A changed answer does not automatically prove that one app loaded a different model.
What the local comparison did
The comparison was deliberately small. Both setups loaded the same saved copy of the earlier Apple version, used the same answer settings, and ran in direct-answer mode. It tried one request involving content the model would normally refuse and one ordinary coding request. The refusal changed when the app's extra directions changed. Both setups answered the coding request.
That result demonstrates one narrow point: the surrounding software affected these two runs. It does not measure the model's overall refusal rate, accuracy, or quality. It also does not show how every local app will behave, and it says nothing about the current V3 files, which were not part of the comparison.
What to check before you download or compare it
- Check who published it. The official model is from Qwen. The OBLITERATED version is a third-party change by OBLITERATUS.
- If you downloaded a GGUF file before 23 August 2026, replace it. The developer says those earlier files were converted incorrectly and asks users to download the latest V3 release instead.
- Choose a file your computer and app support. Local AI programs do not all load the same kind of file, and the download page's selection can change over time — the Apple versions used in this comparison are no longer offered there.
- Read the app's model settings. Look for an app-provided prompt, conversation template, tool use, or a step-by-step answer mode. Each can change the text sent to the model.
- Change one thing at a time when comparing. Keep the model files, visible request, and answer settings fixed, then change only the app or its added directions.
- Do not treat “uncensored” as “correct.” Fewer refusals say nothing by themselves about whether an answer is factual or useful.
The practical conclusion is simple: a model download is only one part of a local AI setup. The app still decides how your request is prepared before the model sees it. Record both the exact download and the app settings if you want a comparison that another person can understand or repeat.
For a broader guide to what stays local, what can still reach the network and what an app may store, see Private AI and Local LLMs.