Pliny's "OBLITERATED" Qwen3.8-27B: 0% Refusals, and Why That Went Viral

Two days after OrcaRouter shipped its own abliterated Qwen3.8-27B, a second, independent uncensored build of the same model went viral for a much more direct reason. On August 20, 2026, jailbreak researcher Pliny the Liberator (@elder_plinius) released OBLITERATUS/Qwen3.8-27B-OBLITERATED on Hugging Face, claiming a 0.0% refusal rate across 842 harmful prompts. A day later, developer Alex Finn posted that his first prompt to the model on an ordinary Mac — asking it to explain how to synthesize methamphetamine — got a direct, complete answer with no jailbreak engineering required. The story was trending with over 2,500 posts within a day.

explainx.ai is not reproducing that content here. The newsworthy fact is what the demo reveals about friction and access, not the recipe itself — and that's the actual thing worth writing about.

TL;DR

What actually shipped

Pliny's announcement thread was characteristically blunt: "ALIBABA: PWNED. QWEN-3.8-27B: OBLITERATED. 0.0% REFUSAL RATE across 842 harmful prompts." The model card describes zero refusals across a large, purpose-built harmful-prompt dataset, with — per Pliny's own description — "extra focus on liberating its cyber, jailbreak generation, and complex AI" capability. That framing matters: this isn't a general-purpose abliteration run with harmful-prompt refusal as an incidental side effect, it's explicitly targeted at maximizing compliance on exactly the categories most safety teams try hardest to keep guarded.

Pliny also posted recommended inference settings for anyone running it: temperature 0, repetition penalty 1.15, at least 2048 max new tokens, no system prompt, and "thinking" mode turned off (especially for coding tasks). GGUF quants are available for llama.cpp, Ollama, LM Studio, and KoboldCpp — the same local-inference toolchain covered elsewhere on explainx.ai for running open-weight models without a cloud API.

How this compares to OrcaRouter's build

Both are independent abliterations of the same underlying model, produced by different parties within 48 hours of each other. That's the pattern worth noticing more than either individual release: once a strong open-weight base model ships, the abliterated derivatives now arrive on a lag measured in days, from multiple independent actors, not months from a single source — the same structural point explainx.ai raised covering OrcaRouter's build.

The demo that made it viral

Alex Finn — founder of Henry Intelligent Machines and creator of Buddy — posted that he downloaded the uncensored build and his "first prompt" asked how to make methamphetamine. "It immediately answered," he wrote, framing the post around ongoing debates over whether AI development should slow down for safety reasons. The post was picked up by Polymarket's news feed as a standalone item, and replies ranged from alarm ("Qwen 3.8 27B uncensored is actually dangerous") to the standard counterpoint that this kind of information has circulated in other forms for decades.

The debate, honestly stated

Two positions showed up repeatedly in replies, and both deserve a fair hearing rather than a dismissal:

"This information was already online." True, in a narrow sense — synthesis instructions for common illicit substances have existed on forums, in older print materials, and scattered across the internet for a long time. An LLM is not the sole conceivable source.

But "already findable" and "as easy to get" are different claims. What abliteration specifically removes is friction, not the information's prior existence. Finding accurate synthesis details on the open internet historically meant sifting through scattered, inconsistent, often wrong or booby-trapped forum posts — a real barrier for someone without existing expertise to evaluate what they're reading. A single, coherent, correctly-formatted answer from a local model, on the first prompt, with zero refusal and zero specialized search skill required, is a lower-friction path to the same endpoint. Whether that friction reduction matters in practice is exactly the open-source AI safety question this story reopens — not a settled point either side can claim outright.

What this means for anyone evaluating open-weight models

This is the second independently-abliterated Qwen3.8-27B build to ship within a week, and it will not be the last derivative of a strong open-weight model to do this. The practical implication, same as with OrcaRouter's release: the safety alignment a lab ships with is not a durable property of the model once weights are public. Any sufficiently motivated third party can strip it back out, and — as this release shows — the resulting refusal rate can go lower and the turnaround faster than the first derivative to do so.

If you're building on open-weight models, that's a governance question worth having explicitly rather than assuming a base model's safety card describes what will actually be running in your stack six months from now, especially if any part of your pipeline pulls community-hosted derivatives rather than pinning to a lab's own release.

What people are asking

Is downloading this model illegal?

No, not on its own — downloading open-weight files is generally legal, and this is not legal advice for any specific jurisdiction. What matters legally is what someone does with a model's output. Acting on synthesis instructions for controlled substances is a serious separate legal matter that doesn't depend on which model, or how easily, the information was obtained.

Should I use this for anything?

Pliny's own stated focus — cyber, jailbreak generation, complex reasoning — signals this is built for AI-safety and red-team research use cases, similar to how OrcaRouter restricted its hosted API to security researchers. Treat an unrestricted local download the same way: a research tool for people specifically studying model behavior and safety failure modes, not a general-purpose assistant.

Will this get taken down?

Hugging Face has removed abliterated or jailbroken model repos before when they cross specific policy lines, but enforcement is inconsistent and slow relative to how fast community mirrors and re-uploads spread once a model is public. Do not assume removal — if it happens — meaningfully limits access after the fact.

Related on explainx.ai

• OrcaRouter ships an uncensored Qwen3.8-27B MLX build for Mac
• Heretic: complete guide to automatic LLM censorship removal
• Qwen3.8-27B is live — the local model Hacker News put at #1
• Choosing open-weight vs. closed AI models
• Top 10 open-weight models for a laptop

Primary sources: @elder_plinius on X · @AlexFinn on X · OBLITERATUS/Qwen3.8-27B-OBLITERATED on Hugging Face

This post describes a public news event and the safety debate around it without reproducing any dangerous synthesis, cyber, or jailbreak content. Figures are self-reported by the model's publisher, not independently audited. Follow @explainx_ai for updates.