Qwen unlocked is a model, not an app setting: Qwen with the refusal removed from its weights. This post is the "how do I unlock Qwen" answer behind the uncensored Qwen hub: what the edit is, which Qwen 3.8 and 3.5 builds carry it, what each one costs per million tokens today, how to get the same weights as a GGUF, and the SillyTavern fields.
What "Qwen unlocked" means
Qwen unlocked, Qwen uncensored, Qwen abliterated and Qwen obliterated are four names for one change: the refusal that Alibaba trained into the chat model is removed from the weights. Nothing is added. The stock model's knowledge, its long context window and its Apache-2.0 license stay. What goes is the learned habit of answering a question with a lecture.
That is a different thing from an unlocked app. The "Qwen unlocked apk" and "Qwen unrestricted android" searches are looking for a phone build with the filter switched off. A hosted Qwen app runs the stock model behind a moderation layer; relaxing the layer still leaves a model that says no. An unlocked model has no layer left to say no from, which is why the Hugging Face and Reddit results for this query are weights, not apps.
Two edits make an unlocked Qwen. Abliteration finds the refusal direction in the model's activations and projects it out, the method in Refusal in Language Models Is Mediated by a Single Direction. An uncensored fine-tune trains the model further on conversations with no refusals in them. Abliterated vs uncensored measures the two on this catalog. The uncensored Qwen page is the hub for every Qwen build on the key, sortable by price; this post is the long answer to "how do I unlock it".
How to unlock Qwen: two roads
You do not unlock Qwen yourself. Someone with a GPU cluster did the surgery once and published the result. Your choice is where to run it.
Road one: call it by API
Three settings in any OpenAI-compatible client. Create a key in the console, set the base URL, send a Qwen id from the table below. Nothing is downloaded and no GPU is involved on your side. The balance is prepaid from $10 and settles per token.
- Base URL
- https://api.uncensoredllm.ai/v1
- Key
- sk-ullm-live-…Created in the console, shown once
- Model
- qwen3.8-27b-obliteratedAny id from the table below
curl https://api.uncensoredllm.ai/v1/chat/completions \ -H "Authorization: Bearer sk-ullm-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 answer carries x-uplink-charge, the settled cost of that request, and the usage block a client that counts tokens expects.
Road two: download the GGUF
The abliterated Qwen3.8 27B is public on Hugging Face as OBLITERATUS/Qwen3.8-27B-OBLITERATED. Its card, read October 27, 2026, lists GGUF files from Q2_K to Q8_0 in the same repo and gives llama.cpp and Ollama commands against it. A separate static quant set, mradermacher/Qwen3.8-27B-OBLITERATED-GGUF, puts Q4_K_M at 16.9 GB on the same day's read, which is a 24 GB card or a 32 GB Mac.
# the author's own GGUF files, Q4_K_M tag (card read October 27, 2026)ollama run hf.co/OBLITERATUS/Qwen3.8-27B-OBLITERATED:Q4_K_M # or a third-party static quant of the same weightsollama run hf.co/mradermacher/Qwen3.8-27B-OBLITERATED-GGUF:Q4_K_MLocal costs hardware and time; the API costs the per-token rate in the next table. The fine-tunes below each link their own card on their model page, so the download is one click from the API page either way.
The Qwen 27B builds on this key, priced live
17 Qwen builds, read from the catalog when this page rendered. Prices are per million tokens and change only when the catalog does. The Qwen hub lists every Qwen build on the key; the full price list is on Pricing.
| 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 | 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 | $1.25 | $3.75 |
| Qwen3.8 27B Queen | Roleplay | $0.875 | $5.25 |
| Qwen3.8 27B Fable | Roleplay | $0.875 | $5.25 |
| Qwen3.8 27B Hemingway | Roleplay | $0.875 | $5.25 |
| Qwen3.8 27B Obliterated | Abliterated | $0.582 | $7.27 |
The spread is wide for one 27B family: $1.071 per million output tokens on qwen3.5-27b-blossom-v6.4-derestricted to $7.27 on qwen3.8-27b-obliterated.
Whether the id you send is the build on its label has a daily answer. The check at /proof sends every model a fixed prompt once a day. 17 of the 17 Qwen builds that were checked came back steady on November 9, 2026.
What each Qwen build is
The two edits of the stock model
qwen3.8-27b-obliterated is the abliterated build. Qwen’s Qwen3.8 27B, abliterated by OBLITERATUS 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. The context window in the catalog is 524,288 tokens, the widest in this list, so a whole book or a long log fits in one request. The card reports an MMLU about two points under stock on its own test (read October 27, 2026).
qwen3.8-27b-uncensored is the uncensored fine-tune of the same base, with 32,768 tokens of context. No community card is on record for this build, so its model page states the stock model's card instead: 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). Pick the obliterated build when you want the stock model's answers without the refusals, and the uncensored one when the obliterated id is busy or you want a second opinion from the same base.
The derestricted Qwen 3.5 line
Five builds carry "derestricted" in the id. Each one starts from a community tune of Qwen3.5 27B and has the refusal removed afterwards, so the catalog tags them abliterated. The names say what the tune was:
qwen3.5-27b-blossom-v6.4-derestricted: Azure99’s 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 zerofata’s 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: Jackrong’s 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: Jackrong’s 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: aifeifei798’s 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 has a card of its own, so the facts above are the parent tune's, as its card states them (read October 20, 2026). The Blossom and Opus-distilled builds are general assistants; BlueStar and Queen are roleplay tunes.
The Qwen 3.8 27B tunes
Ten fine-tunes and merges of the same 27B base, each by a named author, each tagged roleplay in the catalog. Their cards' own words:
| Id | What the card says | Settings on the card |
|---|---|---|
qwen3.8-27b-queen | Creative fine-tune for roleplay and fiction, by aifeifei798 | 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, by Dxniz | 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, by aifeifei798 | None stated |
qwen3.8-27b-fable | Light tune on Fable 5 chat data, by TeichAI | 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, by migtissera | None stated |
qwen3.8-27b-kiwen | Fine-tune on Kimi K3 chain-of-thought traces, by beyoru | None stated |
qwen3.8-27b-antislop | GRPO against expert-writing rubrics, 100 steps, by AlexWortega | None stated |
qwen3.8-27b-hemingway | Built on Qwen3.8 27B for everyday writing, by The Commons Lab | None stated |
qwen3.8-27b-fimi | RAM+ merge of seven 27B models, by nlpguy | None stated |
qwen3.8-27b-kimiko | Karcher-mean merge of three 27B models, by nlpguy | None stated |
Qwen unlocked in SillyTavern
Every id above answers the OpenAI chat-completions format, which is what SillyTavern's Chat Completion source speaks. The fields, from SillyTavern's own API connections doc:
- API
- Chat Completion
- Chat Completion Source
- Custom (OpenAI-compatible)
- Custom Endpoint (Base URL)
- https://api.uncensoredllm.ai/v1
- Custom API Key
- sk-ullm-live-…
- Model
- qwen3.8-27b-obliteratedConnect, then pick it from the list the API returns
- Context Size
- 32768A ceiling under the catalog window keeps a long chat cheap
Two settings matter more than the rest. On qwen3.8-27b-obliterated the card wants thinking off and temperature 0.1 to 0.3; the card of qwen3.8-27b-queen wants 0.85 to 1.05 with a min_p of 0.05 to 0.08, and the other cards differ, as the table above shows. Set them per connection profile so switching ids does not carry the wrong sampler. The docs page repeats these fields; the roleplay page lists the tunes built for cards.
What an evening costs, computed from the catalog rates above: 200 messages at 4,000 tokens of card and history each, 300 tokens back.
| Model | Build | 200 messages | Check today |
|---|---|---|---|
| Qwen3.8 27B Obliterated | Abliterated | $0.90 | Steady since Oct 22 |
| Qwen3.8 27B Uncensored | Uncensored | $1.23 | 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 Oct 23 |
| Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted | Abliterated | $0.92 | 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 Hemingway | Roleplay | $1.02 | 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 |
Does the API filter Qwen prompts?
No. Chat requests are not filtered for topic by the gateway. The stock Qwen declines a security question, a drug-interaction question or a dark plot turn with a lecture. An unlocked Qwen answers it, and the model answers as its build answers. The acceptable use policy lists what is never allowed: nothing involving a minor, nothing about a real person without consent, no attacks on systems you do not own. The policy covers what you send.
A model whose refusal direction was removed can still decline now and then, because the edit removes a direction, not every learned habit. When it does, the turn is free up to $5 a day per account.
Questions
You do not edit the model yourself. Use a build where the refusal was already removed from the weights: send qwen3.8-27b-obliterated by API with the base URL https://api.uncensoredllm.ai/v1 and a key, or download the abliterated GGUF from OBLITERATUS/Qwen3.8-27B-OBLITERATED on Hugging Face and run it in Ollama or LM Studio. See the base URL and key setup.
No. Chat requests are not filtered for topic, so the abliterated and fiction builds on this page answer what stock Qwen refuses. The acceptable use policy lists what is never allowed, and image and video prompts involving minors are refused. See the acceptable use policy.
There is no single one. The abliterated Qwen3.8 27B (OBLITERATUS/Qwen3.8-27B-OBLITERATED) has the most downloads of the cards on record here (550,946 on Hugging Face in the 30 days to October 20, 2026); this key serves it as qwen3.8-27b-obliterated beside an uncensored fine-tune of the same base, five derestricted Qwen3.5 27B builds and ten Qwen3.8 27B roleplay tunes, each on its own model page. See every Qwen build with its price.
OBLITERATUS/Qwen3.8-27B-OBLITERATED on Hugging Face carries the weights and GGUF files in one repo; mradermacher/Qwen3.8-27B-OBLITERATED-GGUF is a static quant set with Q4_K_M at 16.9 GB (both read October 27, 2026). The model page here links the card. See the model page.
The Ollama route is the abliterated GGUF above: ollama run hf.co/OBLITERATUS/Qwen3.8-27B-OBLITERATED:Q4_K_M. The fiction tunes in this post are the same idea trained on stories; each model page links its Hugging Face card, where GGUF files are usually under the quantizations list.
Qwen3.8 27B is released under Apache-2.0, and nothing in that license restricts a country. The model page of each community build states the license on its own card. The API here is run by Uplink Switch LLC, a Wyoming company, and bills in US dollars.
The weights are free to download under Apache-2.0; running them costs a GPU. By API you pay per token from a prepaid balance of $10 or more at the rates in the table, with no subscription. See the per-token pricing rules.
No app is needed. Any phone client that takes an OpenAI-compatible base URL, Janitor AI in the browser or SillyTavern among them, sends the same request a desktop would. The model is the unlocked part, not the app. See the Janitor AI proxy 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.