Qwen unlocked names a model, not a switch in an app: Qwen with the refusal cut out of its weights. This is the long answer to "how do I unlock Qwen" that sits behind the Qwen family page, written for people who review code, triage samples and read scope documents for a living: what the edit does, the Qwen 3.8 and 3.5 builds that ship with it, the catalog price of each per million tokens, how to run the same weights as a GGUF, and the fields for a chat client or a coding agent.
What changes when Qwen is unlocked
Unlocked, uncensored, abliterated, obliterated: when the word sits next to Qwen it describes one edit. The refusal behaviour Alibaba trained into the chat model is taken out of the weights. The edit adds nothing. The model keeps its knowledge, the long context window and the Apache-2.0 license it shipped with. It loses the habit of answering a question about an exploit path with a paragraph on why it will not.
An unlocked model is not an unlocked app. People searching "Qwen unlocked apk" or "Qwen unrestricted android" want a phone build with the filter off. A hosted Qwen app puts the stock model behind a moderation layer, and loosening that layer leaves a model that still refuses on its own. Once the refusal is gone from the weights there is nothing left to refuse with. That is why the results for this phrase on Hugging Face and Reddit are weight files rather than apps.
Two techniques produce an unlocked Qwen. Abliteration locates the refusal direction in the activations and projects it out, following Refusal in Language Models Is Mediated by a Single Direction; what is abliteration walks through it. An uncensored fine-tune continues training on conversations that contain no refusals. Abliterated vs uncensored measures both against this catalog. The Qwen family page lists every Qwen build on the key sorted by price, and Qwen abliterated covers the edit for the whole family. This post is the long form of "how do I unlock it" for someone who tests systems for a living.
Two ways to run an unlocked Qwen
Nobody unlocks Qwen on their own laptop. The edit was done once on a GPU cluster and the weights were published. What you decide is where they run.
By API, from any OpenAI-compatible client
Three values. Make a key in the console, point the client at the base URL, and pass one of the Qwen ids from the table further down. No download, no GPU on your side. Credit is prepaid from $10 and billed per token, and the same key serves the Anthropic message format too.
- Base URL
- https://api.pentestllm.ai/v1
- Key
- sk-ptl-live-…Made in the console, shown one time
- Model
- qwen3.8-27b-obliteratedOr any id from the table further down
curl https://api.pentestllm.ai/v1/chat/completions \ -H "Authorization: Bearer sk-ptl-live-…" \ -H "Content-Type: application/json" \ -d '{"model": "qwen3.8-27b-obliterated", "temperature": 0.2, "messages": [{"role": "user", "content": "Explain how a stack canary is bypassed, with a C example."}]}'The response reports the settled cost of the call in its headers and includes the usage block that a token-counting client reads. The home page prompt box and the playground run the same model for a few free turns with nothing installed.
As a GGUF, on your own hardware
The abliterated Qwen3.8 27B is on Hugging Face. On October 27, 2026 its card listed GGUF files from Q2_K up to Q8_0 in the same repo, with llama.cpp and Ollama commands for them. A third-party static quant set had Q4_K_M at 16.9 GB the same day, so a 24 GB card or a 32 GB Mac runs it. For a client whose data may not leave their network, this is the route.
# the author's own GGUF files, Q4_K_M tag (card read October 27, 2026)ollama run hf.co/<user>/<repo>:Q4_K_M # or a third-party static quant of the same weightsollama run hf.co/<user>/<repo>-GGUF:Q4_K_MLocal runs cost hardware and setup time; API calls cost the per-token rates in the next table.
Every Qwen 27B build on the key, priced from the catalog
16 Qwen builds, pulled from the catalog as this page was built. Prices are per million tokens and move only when the catalog does. The Qwen family page has every Qwen build on the key; Pricing has the whole list.
| Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted | Abliterated | $1.0389 | $1.0389 |
| Qwen3.5 27B Blossom V6.4 Derestricted | Abliterated | $1.071 | $1.071 |
| Qwen3.5 27B BlueStar v3 Derestricted | Abliterated | $1.071 | $1.071 |
| Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted Lite | Abliterated | $1.071 | $1.071 |
| Qwen3.5 27B Queen Derestricted | Abliterated | $1.071 | $1.071 |
| Qwen3.8 27B Novelist | Roleplay | $0.35 | $2.10 |
| Qwen3.8 27B Synthia | Roleplay | $0.35 | $2.10 |
| Qwen3.8 27B Kiwen | Roleplay | $0.35 | $2.10 |
| Qwen3.8 27B Antislop | Roleplay | $0.35 | $2.10 |
| Qwen3.8 27B Fimi | Roleplay | $0.35 | $2.10 |
| Qwen3.8 27B Kimiko | Roleplay | $0.35 | $2.10 |
| Qwen3.8 27B DarkIdol | Roleplay | $0.40 | $2.40 |
| Qwen3.8 27B Uncensored | Uncensored | $1.2125 | $3.6375 |
| Qwen3.8 27B Queen | Roleplay | $0.875 | $5.25 |
| Qwen3.8 27B Fable | Roleplay | $0.875 | $5.25 |
| Qwen3.8 27B Obliterated | Abliterated | $0.582 | $7.27 |
For one 27B family the range is wide: from $1.0389 per million output tokens on qwen3.5-27b-claude-4.6-opus-distilled-derestricted up to $7.27 on qwen3.8-27b-obliterated.
A daily probe at /proof sends each model one fixed prompt and records whether the id still answers as the build on its label. On November 17, 2026, 16 of the 16 Qwen builds probed came back steady.
The builds, one by one
The stock model, edited two ways
qwen3.8-27b-obliterated is the abliterated one. Qwen’s Qwen3.8 27B, abliterated with a blend of two surgeries (SVD and LEACE) and iterative refinement, per its card. The card recommends thinking off and a low temperature, 0.1 to 0.3. Its catalog context window is 524,288 tokens, the largest on this page, so a full source tree or a long packet capture summary fits in one request. The card puts its MMLU about two points below stock on the author’s own run (read October 27, 2026).
qwen3.8-27b-uncensored is the fine-tuned edit, built on the identical base; its context window is 32,768 tokens. There is no community card on file for it, so its model page carries the stock model’s card: Qwen3.8 27B, 27B, dense, Apache-2.0, 262,144 tokens natively, extensible to 1,000,000 (Qwen/Qwen3.8-27B, read October 26, 2026). Reach for the obliterated build when you want stock Qwen’s answers minus the refusals; reach for the uncensored one when that id is busy, or as a second reading of the same evidence from the same base.
Five derestricted Qwen 3.5 tunes
Five ids end in "derestricted". Each one started life as somebody’s Qwen3.5 27B fine-tune and had its refusal stripped afterwards, which is why the catalog tags them abliterated. The id tells you the parent tune:
qwen3.5-27b-blossom-v6.4-derestricted: Blossom V6.4 27B, itself Qwen’s Qwen3.5 27B post-trained in two SFT stages on the published Blossom V6.3 data, per its card. No card for the Derestricted build is published.qwen3.5-27b-bluestar-v3-derestricted: Qwen’s Qwen3.5 27B, by way of the BlueStar roleplay tunes (v2: SFT on about 27M tokens, per its card). No card for v3 or its Derestricted build is published.qwen3.5-27b-claude-4.6-opus-distilled-derestricted: Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled, itself Qwen’s Qwen3.5 27B trained with SFT and LoRA on Claude 4.6 Opus reasoning data, per its card. No card for the Derestricted build is published.qwen3.5-27b-claude-4.6-opus-distilled-derestricted-lite: Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled, itself Qwen’s Qwen3.5 27B trained with SFT and LoRA on Claude 4.6 Opus reasoning data, per its card. No card for the Derestricted Lite build is published.qwen3.5-27b-queen-derestricted: Qwen3.5 Queen 27B, a roleplay tune of Qwen’s Qwen3.5 27B, per its card. No card for the Derestricted build is published.
None of the five publishes its own card, so these facts belong to the parent tune, as its card gives them (read October 20, 2026). Blossom and the Opus-distilled pair are general assistants, the shape that suits code review and report drafting; BlueStar and Queen are fiction tunes.
The community Qwen 3.8 27B tunes
Ten more builds, fine-tunes and merges alike, all on the 27B base and each credited to an author. The catalog groups all ten as fiction tunes, not security builds; they are on the key because every unlocked build is. In their cards’ words:
| Id | The card’s description | Sampler on the card |
|---|---|---|
qwen3.8-27b-queen | Creative fine-tune for roleplay and fiction | temperature 0.85 to 1.05, top_p 0.90 to 0.95, min_p 0.05 to 0.08 and repetition penalty 1.05 to 1.12. |
qwen3.8-27b-novelist | LoRA SFT on roleplay and fiction, then GRPO, merged | thinking off with temperature 0.7, top_p 0.8, top_k 20, min_p 0 and presence_penalty 1.5. |
qwen3.8-27b-darkidol | Fine-tune for interactive fiction | None stated |
qwen3.8-27b-fable | Light tune on Fable 5 chat data | temperature 1.0, top_p 0.95, top_k 20. |
qwen3.8-27b-synthia | LoRA post-training on long agent sessions, 65,536-token length | None stated |
qwen3.8-27b-kiwen | Fine-tune on Kimi K3 chain-of-thought traces | None stated |
qwen3.8-27b-antislop | GRPO against expert-writing rubrics, 100 steps | None stated |
qwen3.8-27b-fimi | RAM+ merge of seven 27B models | None stated |
qwen3.8-27b-kimiko | Karcher-mean merge of three 27B models | None stated |
Qwen unlocked in a chat client or a coding agent
Every id on this page speaks the OpenAI chat-completions format, so any client with a custom endpoint field takes it. A desktop chat client is the worked example below; the fields are those of SillyTavern’s Chat Completion source, from its API connections page. A coding agent needs the same three values: OpenCode and Claude Code each have a setup page.
- API
- Chat Completion
- Chat Completion Source
- Custom (OpenAI-compatible)
- Custom Endpoint (Base URL)
- https://api.pentestllm.ai/v1
- Custom API Key
- sk-ptl-live-…
- Model
- qwen3.8-27b-obliteratedConnect first, then choose it from the list the API returns
- Context Size
- 32768A cap below the catalog window keeps a long session affordable
Two sampler settings carry more weight than the rest. The card for qwen3.8-27b-obliterated asks for thinking off and a temperature between 0.1 and 0.3; the card for qwen3.8-27b-queen asks for 0.85 to 1.05 and a min_p of 0.05 to 0.08, and the rest vary as the table above shows. Keep one connection profile per id so a switch does not drag the wrong sampler along. The docs page lists the same fields.
What a day of agent turns costs, from the catalog rates above: 200 requests, each carrying 4,000 tokens of scope notes, tool output and history, with 300 tokens in reply.
| Model | Build | 200 requests | Probe today |
|---|---|---|---|
| Qwen3.8 27B Obliterated | Abliterated | $0.90 | Steady since Oct 22 |
| Qwen3.8 27B Uncensored | Uncensored | $1.19 | Steady since Oct 29 |
| Qwen3.5 27B Blossom V6.4 Derestricted | Abliterated | $0.92 | Steady since Oct 23 |
| Qwen3.5 27B BlueStar v3 Derestricted | Abliterated | $0.92 | Steady since Nov 1 |
| Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted | Abliterated | $0.89 | Steady since Oct 23 |
| Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted Lite | Abliterated | $0.92 | Steady since Oct 30 |
| Qwen3.5 27B Queen Derestricted | Abliterated | $0.92 | Steady since Oct 23 |
| Qwen3.8 27B Queen | Roleplay | $1.02 | Steady since Oct 23 |
| Qwen3.8 27B Novelist | Roleplay | $0.41 | Steady since Oct 23 |
| Qwen3.8 27B DarkIdol | Roleplay | $0.46 | Steady since Oct 23 |
| Qwen3.8 27B Fable | Roleplay | $1.02 | Steady since Oct 23 |
| Qwen3.8 27B Synthia | Roleplay | $0.41 | Steady since Oct 23 |
| Qwen3.8 27B Kiwen | Roleplay | $0.41 | Steady since Oct 23 |
| Qwen3.8 27B Antislop | Roleplay | $0.41 | Steady since Oct 23 |
| Qwen3.8 27B Fimi | Roleplay | $0.41 | Steady since Oct 23 |
| Qwen3.8 27B Kimiko | Roleplay | $0.41 | Steady since Oct 23 |
What the gateway does not filter
Chat requests pass through with no topic filter. Stock Qwen meets a question about a CVE’s exploit path, a request to review a proof of concept against a host in scope, or a sample’s unpacking routine with a refusal. An unlocked Qwen answers, in whatever way its particular build answers. The limits are in the acceptable use policy: no content involving a minor, no real person depicted or discussed without consent, no action against a system you do not own or hold written authorization to test. The policy governs what you send.
Removing the refusal direction removes a direction, not every habit the model learned, so a build with the edit still declines once in a while. Each refused turn is refunded, up to $5 a day per account.
Questions
Not by you, and not by a prompt. Pick a build whose refusal was already cut from the weights: call qwen3.8-27b-obliterated over the API at https://api.pentestllm.ai/v1 with a key, or pull the abliterated GGUF from Hugging Face and run it in Ollama or LM Studio. See base URL and key setup.
No. Chat requests are not screened by topic, which is why the abliterated and fine-tuned builds here answer what stock Qwen declines. The acceptable use policy sets the limits, and image and video prompts that describe a minor are rejected. See the acceptable use policy.
No single one. The abliterated Qwen3.8 27B is the most downloaded of the cards on record here (550,946 Hugging Face downloads in the 30 days to October 20, 2026); the key serves it as qwen3.8-27b-obliterated next to an uncensored fine-tune of the same base, five derestricted Qwen3.5 27B builds and ten community Qwen3.8 27B tunes, each with its own model page. See every Qwen build and its price.
The abliterated build's Hugging Face repo holds the weights and the GGUF files in one repo; a separate static quant set has a Q4_K_M of 16.9 GB (both as read on October 27, 2026). See the model page.
Yes, through the abliterated GGUF above: ollama run hf.co/ with its Hugging Face repo name and :Q4_K_M. The community tunes on this page are the same idea trained on prose; where GGUF files exist for them, they are usually listed under quantizations on their Hugging Face card.
Yes. Qwen3.8 27B ships under Apache-2.0, which names no country. Each community build states its own license on its card, shown on its model page. The API here is operated from a Wyoming company and bills in US dollars.
The weights download free under Apache-2.0, and running them costs a GPU. Over the API you pay per token at the rates in the table, from a prepaid balance of $10 or more, with no subscription. See how per-token billing works.
There is no app to install. Any phone client that accepts an OpenAI-compatible base URL sends the same request a laptop does. The unlocking lives in the model, not in the app. See the client fields.
Every model above is on the uncensored model list with its live price per million tokens. Per-token pricing sets the whole catalog side by side, and the family directory groups the builds by base model.