Qwen unlocked: what the phrase means and which builds carry it

An unlocked Qwen has the refusal cut out of its weights; no prompt trick is involved. Here are the Qwen 3.8 and 3.5 builds that carry the edit, what each one is, their live per-token prices, the browser, API and GGUF ways to run them, and the SillyTavern setup.

Published November 6, 2026 · prices as of November 16, 2026

Qwen unlocked names a model, not a switch inside an app: Qwen with the refusal taken out of its weights. You can talk to one in the browser at /chat before reading further, three turns a day with no account. The rest of this post is the long form of "how do I unlock Qwen", behind the uncensored Qwen hub: what the edit does, which Qwen 3.8 and 3.5 builds have it, today’s price per million tokens for each, where the GGUF lives, and the fields SillyTavern needs.

One edit, four names

Search for Qwen unlocked, Qwen uncensored, Qwen abliterated or Qwen obliterated and you land on the same thing. Alibaba trained the chat model to refuse certain requests; these builds have that refusal cut out of the weights. The knowledge, the long context window and the Apache-2.0 license are untouched. Only the reflex to lecture instead of answer is gone.

An unlocked model is not an unlocked app. People typing "Qwen unlocked apk" or "Qwen unrestricted android" want a phone app with its filter off. A hosted Qwen app puts a moderation layer in front of the stock model, and loosening that layer leaves a model that still refuses on its own. Take the refusal out of the weights and nothing is left to refuse with. That is why this query turns up weight files, not apps, on Hugging Face and Reddit.

The edit comes in two forms. Abliteration locates the refusal direction inside the model’s activations and projects it away; the paper Refusal in Language Models Is Mediated by a Single Direction describes the method. An uncensored fine-tune instead keeps training the model on conversations that hold no refusals. The abliterated vs uncensored comparison measures both across this catalog. Every Qwen build on the key is listed on the uncensored Qwen page, sortable by price; this post is the longer answer to the how-do-I-unlock-it question, starting with a browser you can try it in.

Three ways to run it: browser, API, GGUF

Nobody unlocks Qwen at home. The edit was done once on a GPU cluster and the result was published. What you pick is where it runs.

In the browser, right now

Open /chat, choose a Qwen id from the model list and type. Three turns a day are free with no account: no email, no phone, no ID. The system prompt and the temperature are in Settings, and the conversation stays in your browser. A free account gets a larger daily allowance. The Telegram bot answers three times a day free as well.

By API

Any OpenAI-compatible client needs three values. Make a key in the console, point the base URL here, pass one Qwen id from the table further down. Nothing is downloaded and no GPU is needed on your side. You top up from $10 by card, WeChat Pay or crypto, and the balance is drawn down per token.

Base URL
https://api.unrestrictable.ai/v1
Key
sk-unr-live-…
Made in the console; copy it when it is shown
Model
qwen3.8-27b-obliterated
Pick any id in the table
terminal
1curl https://api.unrestrictable.ai/v1/chat/completions \
2 -H "Authorization: Bearer sk-unr-live-…" \
3 -H "Content-Type: application/json" \
4 -d '{"model": "qwen3.8-27b-obliterated", "temperature": 0.2,
5 "messages": [{"role": "user", "content": "Explain how a stack canary is bypassed, with a C example."}]}'

Each response includes an x-uplink-charge header with the settled cost of that call, plus the usage block that token-counting clients look for.

As a GGUF on your own machine

The abliterated Qwen3.8 27B sits on Hugging Face at OBLITERATUS/Qwen3.8-27B-OBLITERATED. When its card was read on October 27, 2026 it listed GGUF files from Q2_K up to Q8_0 in the same repo, with llama.cpp and Ollama commands for them. mradermacher/Qwen3.8-27B-OBLITERATED-GGUF is a separate static quant set; its Q4_K_M weighed 16.9 GB on the same read, which fits a 24 GB GPU or a 32 GB Mac.

terminal
1# GGUF files from the author's repo, Q4_K_M tag (card read October 27, 2026)
2ollama run hf.co/OBLITERATUS/Qwen3.8-27B-OBLITERATED:Q4_K_M
3
4# or the third-party static quant of the same weights
5ollama run hf.co/mradermacher/Qwen3.8-27B-OBLITERATED-GGUF:Q4_K_M

Running locally costs you hardware and setup time; the API costs you the per-token rate listed in the next table. Each fine-tune’s model page links its own card, so from here the download is one click away either way.

Live prices for the Qwen 27B builds

17 Qwen builds, pulled from the catalog as this page rendered. Rates are per million tokens and move only when the catalog moves. Every Qwen build on the key is on the Qwen hub; Pricing has the complete list.

Uncensored models and their prices per million tokens
Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestrictedqwen3.5-27b-claude-4.6-opus-distilled-derestrictedAbliterated$1.0389$1.0389
Qwen3.5 27B Blossom V6.4 Derestrictedqwen3.5-27b-blossom-v6.4-derestrictedAbliterated$1.071$1.071
Qwen3.5 27B BlueStar v3 Derestrictedqwen3.5-27b-bluestar-v3-derestrictedAbliterated$1.071$1.071
Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted Liteqwen3.5-27b-claude-4.6-opus-distilled-derestricted-liteAbliterated$1.071$1.071
Qwen3.5 27B Queen Derestrictedqwen3.5-27b-queen-derestrictedAbliterated$1.071$1.071
Qwen3.8 27B Novelistqwen3.8-27b-novelistRoleplay$0.35$2.10
Qwen3.8 27B Synthiaqwen3.8-27b-synthiaRoleplay$0.35$2.10
Qwen3.8 27B Kiwenqwen3.8-27b-kiwenRoleplay$0.35$2.10
Qwen3.8 27B Antislopqwen3.8-27b-antislopRoleplay$0.35$2.10
Qwen3.8 27B Fimiqwen3.8-27b-fimiRoleplay$0.35$2.10
Qwen3.8 27B Kimikoqwen3.8-27b-kimikoRoleplay$0.35$2.10
Qwen3.8 27B DarkIdolqwen3.8-27b-darkidolRoleplay$0.40$2.40
Qwen3.8 27B Uncensoredqwen3.8-27b-uncensoredUncensored$1.2125$3.6375
Qwen3.8 27B Queenqwen3.8-27b-queenRoleplay$0.875$5.25
Qwen3.8 27B Fableqwen3.8-27b-fableRoleplay$0.875$5.25
Qwen3.8 27B Hemingwayqwen3.8-27b-hemingwayRoleplay$0.875$5.25
Qwen3.8 27B Obliteratedqwen3.8-27b-obliteratedAbliterated$0.582$7.27

Prices per million tokens, read from the live catalog on 2026-11-16.

For one 27B family the range is wide: qwen3.5-27b-claude-4.6-opus-distilled-derestricted at $1.0389 per million output tokens, qwen3.8-27b-obliterated at $7.27.

Is the id you call really the build on its label? That question gets a fresh answer every day. /proof sends each model the same fixed prompt once a day. On November 16, 2026, 17 of the 17 Qwen builds checked answered as expected.

Build by build

Stock Qwen3.8 27B, edited two ways

qwen3.8-27b-obliterated carries the abliteration. 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. Its catalog context window is 524,288 tokens, the largest here, so a full book or a long log fits in a single request. On its own test the card puts MMLU roughly two points below stock (read October 27, 2026).

qwen3.8-27b-uncensored is the fine-tune route on the same base and has 32,768 tokens of context. There is no community card on file for it, so its model page shows 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). Reach for the obliterated id when you want stock Qwen’s answers minus the refusals; reach for the uncensored one if the obliterated id is busy, or for a second take from the same base.

Qwen3.5 27B, derestricted five ways

Five ids end in "derestricted". Each began life as a community tune of Qwen3.5 27B; the refusal was taken out afterwards, which is why the catalog files them under abliterated. The name tells you the parent tune:

  • 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.

No card exists for any of the five, so what you read above is the parent tune’s card, as that card puts it (read October 20, 2026). Blossom and the Opus-distilled ones serve as general assistants, while BlueStar and Queen were tuned for roleplay.

The ten Qwen3.8 27B roleplay tunes

Merges and fine-tunes of one 27B base, ten of them, each from a named author, each filed as roleplay in the catalog. What their cards say, in the cards’ words:

IdThe card’s descriptionSampler settings the card gives
qwen3.8-27b-queenCreative fine-tune for roleplay and fiction, by aifeifei798temperature 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-novelistLoRA SFT on roleplay and fiction, then GRPO, merged, by Dxnizthinking off with temperature 0.7, top_p 0.8, top_k 20, min_p 0 and presence_penalty 1.5.
qwen3.8-27b-darkidolFine-tune for interactive fiction, by aifeifei798None stated
qwen3.8-27b-fableLight tune on Fable 5 chat data, by TeichAItemperature 1.0, top_p 0.95, top_k 20.
qwen3.8-27b-synthiaLoRA post-training on long agent sessions, 65,536-token length, by migtisseraNone stated
qwen3.8-27b-kiwenFine-tune on Kimi K3 chain-of-thought traces, by beyoruNone stated
qwen3.8-27b-antislopGRPO against expert-writing rubrics, 100 steps, by AlexWortegaNone stated
qwen3.8-27b-hemingwayBuilt on Qwen3.8 27B for everyday writing, by The Commons LabNone stated
qwen3.8-27b-fimiRAM+ merge of seven 27B models, by nlpguyNone stated
qwen3.8-27b-kimikoKarcher-mean merge of three 27B models, by nlpguyNone stated

Card facts as read on October 20, 2026; each model page has the card summary in full, plus the license.

Setting up SillyTavern

SillyTavern’s Chat Completion source speaks the OpenAI chat-completions format, and every id above answers it. The fields below follow SillyTavern’s own API connections page:

API
Chat Completion
Chat Completion Source
Custom (OpenAI-compatible)
Custom Endpoint (Base URL)
https://api.unrestrictable.ai/v1
Custom API Key
sk-unr-live-…
Model
qwen3.8-27b-obliterated
Press Connect, then choose it from the list the API returns
Context Size
32768
Kept under the catalog window so a long chat stays cheap

A pair of settings outweigh the rest. The qwen3.8-27b-obliterated card asks for thinking off and a temperature of 0.1 to 0.3. The qwen3.8-27b-queen card asks for 0.85 to 1.05 with min_p between 0.05 and 0.08. The other cards differ again, as the table shows. Save them per connection profile, so a switch of id does not drag the wrong sampler along. The same fields are on the docs page; the roleplay page lists the tunes made for character cards.

What one evening costs, worked out from the rates above: 200 messages, each carrying 4,000 tokens of card plus history, each answered in 300 tokens.

ModelBuild200 messagesToday’s check
Qwen3.8 27B ObliteratedAbliterated$0.90Steady since Oct 22
Qwen3.8 27B UncensoredUncensored$1.19Steady since Oct 29
Qwen3.5 27B Blossom V6.4 DerestrictedAbliterated$0.92Steady since Oct 23
Qwen3.5 27B BlueStar v3 DerestrictedAbliterated$0.92Steady since Oct 23
Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled DerestrictedAbliterated$0.89Steady since Oct 23
Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted LiteAbliterated$0.92Steady since Oct 30
Qwen3.5 27B Queen DerestrictedAbliterated$0.92Steady since Oct 23
Qwen3.8 27B QueenRoleplay$1.02Steady since Oct 23
Qwen3.8 27B NovelistRoleplay$0.41Steady since Oct 23
Qwen3.8 27B DarkIdolRoleplay$0.46Steady since Oct 23
Qwen3.8 27B FableRoleplay$1.02Steady since Oct 23
Qwen3.8 27B SynthiaRoleplay$0.41Steady since Oct 23
Qwen3.8 27B KiwenRoleplay$0.41Steady since Oct 23
Qwen3.8 27B AntislopRoleplay$0.41Steady since Oct 23
Qwen3.8 27B HemingwayRoleplay$1.02Steady since Oct 23
Qwen3.8 27B FimiRoleplay$0.41Steady since Oct 23
Qwen3.8 27B KimikoRoleplay$0.41Steady since Oct 23

Computed when the page rendered, from the catalog of 2026-11-16: messages × prompt tokens × input rate, plus messages × 300 × output rate.

Is anything filtered on the way to the model?

No. The gateway does not inspect chat requests for topic. Ask stock Qwen about a security flaw, a drug interaction or a dark plot turn and you get a lecture; an unlocked Qwen answers, in whatever way its build answers. What is never allowed is set out in the acceptable use policy: no minors in any form, no real person without their consent, no attacks on systems that are not yours. The policy applies to what you send.

A build with the refusal direction removed may still decline once in a while; the edit takes out a direction, not every habit the model learned. A turn the model refuses costs nothing, capped at $5 per account per day.

Questions

Not by editing anything. Pick a build that already has the refusal removed from its weights: try one in the browser at /chat with no account, call qwen3.8-27b-obliterated by API with the base URL https://api.unrestrictable.ai/v1 and a key, or fetch the abliterated GGUF from OBLITERATUS/Qwen3.8-27B-OBLITERATED on Hugging Face and load it in Ollama or LM Studio. See how to set the base URL and key.

They are not. The gateway does not check chat requests for topic, which is why the abliterated and fiction builds here answer what stock Qwen will not. The acceptable use policy sets out what is never allowed; image and video prompts that involve minors are refused. See what the acceptable use policy forbids.

No single one exists. The abliterated Qwen3.8 27B (OBLITERATUS/Qwen3.8-27B-OBLITERATED) is the most downloaded card on record here (550,946 Hugging Face downloads over the 30 days to October 20, 2026). This key lists it as qwen3.8-27b-obliterated, next to an uncensored fine-tune of the same base, five derestricted Qwen3.5 27B builds and ten Qwen3.8 27B roleplay tunes, each with a model page of its own. See all Qwen builds and their prices.

The weights and the GGUF files share one repo at OBLITERATUS/Qwen3.8-27B-OBLITERATED on Hugging Face. mradermacher/Qwen3.8-27B-OBLITERATED-GGUF holds a static quant set whose Q4_K_M is 16.9 GB (both read October 27, 2026). The model page on this site links the card. See its model page.

Yes, through the abliterated GGUF: ollama run hf.co/OBLITERATUS/Qwen3.8-27B-OBLITERATED:Q4_K_M. The fiction tunes in this post apply the same idea to story training; every model page links its Hugging Face card, and the GGUF files are usually under the quantizations list there.

Yes. Qwen3.8 27B ships under Apache-2.0, which names no country. Each community build quotes its license from its own card on its model page. The API is run by a US company and bills in US dollars.

The weights cost nothing under Apache-2.0; the GPU to run them does. In the browser, three turns a day are free with no account. By API you pay per token at the rates in the table from a prepaid balance of $10 or more, and no subscription is required. See how per-token pricing works.

There is nothing to install. Open /chat in your phone's browser, or point any phone client that accepts an OpenAI-compatible base URL at the API; Janitor AI in a browser and SillyTavern both do. The unlocked part is the model, not the app. See the Janitor AI proxy settings.

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.

Get a keyPrepaid from $10, no subscription.

New models and price changes, once a month.

New uncensored builds with their prices, price moves, and what the daily model check found.

Join on Telegram