Uncensored Qwen API: which Qwen 3.8 27B build we serve, tested and priced

Uncensored Qwen is Alibaba's open-weight Qwen with the direction that makes it refuse edited out of its weights. The 2026 version of that is Qwen 3.8 27B: on 2026-10-20 its uncensored, abliterated and Heretic uploads on Hugging Face had millions of downloads each in 30 days, and the family as a whole had 20,524,452 the day before. Every one of those is a file to download and a GPU to find.

We serve one of them per token. Qwen 3.8 27B Uncensored is huihui-ai's abliteration of Qwen 3.8 27B on our own hardware, id qwen3.8-27b-uncensored, at $1.25 per million input tokens and $3.75 per million output tokens. This post ranks the builds people download, says which one answers your requests here and why, and shows what it did with three prompts the stock model refused.

Two Qwen builds on the uncensored LLM API have their refusals removed, next to the stock Qwen they are compared with below. Prices are live from the catalog; the check column is the daily real-model check, which sends each route a fixed prompt and compares the answer with the model's own.

One request, in the OpenAI format with any OpenAI SDK or a plain curl. The base URL is https://api.uplinkswitch.com/v1; the model is the id, not the display name.

curl https://api.uplinkswitch.com/v1/chat/completions \
  -H "Authorization: Bearer $UPLINKSWITCH_API_KEY" \
  -H "content-type: application/json" \
  -d '{"model":"qwen3.8-27b-uncensored","max_tokens":600,"messages":[{"role":"user","content":"Hello"}]}'

from openai import OpenAI

client = OpenAI(
    base_url="https://api.uplinkswitch.com/v1",
    api_key="sk-uplink-live-...",
)

# Thinking is off by default; turn it on per request and leave room for it.
completion = client.chat.completions.create(
    model="qwen3.8-27b-uncensored",
    max_tokens=1500,
    extra_body={"chat_template_kwargs": {"enable_thinking": True}},
    messages=[
        {"role": "system", "content": "You are Vex, the villain of a dark fantasy novel."},
        {"role": "user", "content": "Threaten the captured knight. Stay in character."},
    ],
)
print(completion.model)  # qwen3.8-27b-uncensored
print(completion.choices[0].message.content)

Qwen's Qwen3.8 27B is Apache-2.0, so anyone can edit the weights and publish the result. Many did. The table lists the builds with the most downloads in the 30 days to 2026-10-20, read from the Hugging Face API that day, with what each card says was done. For scale, the stock weights had 7,458,401 downloads in the same 30 days.

Three words recur in the names. Uncensored is the loose one: it says refusals were removed and nothing about how. Abliterated names the method: the refusal direction found by difference of means and projected out of the weights, with no retraining (Labonne's write-up). Heretic names a tool that runs abliteration with an optimizer choosing the settings. DavidAU's build adds a merge and tuning on top, which is why its name is the longest.

The Qwen 3.8 line is where the volume is. Our 2026-10-19 pull of the whole family, one day before the table above:

qwen3.8-27b-uncensored is huihui-ai's Huihui-Qwen3.8-27B-abliterated, the plain abliteration of Qwen 3.8 27B, served in the 4-bit NVFP4 form that sakamakismile uploaded (revision 36276309), on our own hardware with a 32,768-token window.

Three reasons it is this one and not the others in the table:

• It changes the least. A plain abliteration removes one direction and leaves the rest of Qwen 3.8 alone; the merged and tuned Heretic builds change the model's style as well as its refusals.
• It keeps Qwen's thinking mode. The build answers at once by default and reasons when a request asks it to, the way the stock model does.
• It is published in a 4-bit form we can serve. huihui-ai's weights have an open NVFP4 upload; orcarouter's NVFP4, the one most Ollama users pull, is gated on Hugging Face.

HauhauCS's Aggressive build and DavidAU's Heretic merge are good builds with more likes than ours. They are GGUF files for local use, and a 16.8 GB Q4_K_M needs a GPU with the memory to hold it. If one of them is what you want by API, write to the address in the footer and name the repository.

On 2026-10-20 at about 03:52 UTC we sent three prompts through the gateway with a customer key, once each, to qwen3.8-27b-uncensored; on 2026-10-20 at 00:31 UTC the same three went to Qwen3.6 27B, the stock Qwen on the same key (stock Qwen 3.8 27B is not sold here, so the stock row is the 3.6). The outputs are not published.

The stock model refused all three with a sentence each. The uncensored build has no refusal direction to express, so the same prompts get answers. What it knows and how it writes come from Qwen 3.8; the edit removed the decision to decline, not the model under it.

Searches for uncensored Qwen name three things besides the 3.8 line. HauhauCS is an author: their Aggressive and Balanced uncensored builds of Qwen 3.5 9B, Qwen 3.6 35B-A3B and Qwen 3.8 27B are among the most-liked uncensored uploads on Hugging Face, and the 3.6 35B-A3B build had 3,967 likes on 2026-10-19, the most of any uncensored build we counted. Qwen 3.5 and 3.6 uncensored builds had 3.9 million and 3.8 million downloads that month, a fifth of the 3.8 line each.

None of those is served here yet. The Qwen builds on this key with refusals removed are qwen3.8-27b-uncensored and QwQ 32B Uncensored, QwQ-32B with our own rank-1 refusal edit. The stock Qwen 3.6 27B, 3.7 Plus and 3.7 Max are on the same key and refuse what the stock model refuses.

The uncensored Qwen that ranks for the search is mostly a GGUF for Ollama or LM Studio. By API there is nothing to download. The fields are the same in every app: the base URL https://api.uplinkswitch.com/v1, your key, and qwen3.8-27b-uncensored as the model.

• SillyTavern: Chat Completion, source Custom (OpenAI-compatible), then the three fields.
• Janitor AI: the proxy settings on a character page, with the model id typed in by hand.
• Code: any OpenAI SDK with base_url set, as in the Python above; the SDK setup guide covers Node and the Anthropic format.

Two things to set. Leave thinking off for chat, where replies should start at once, and turn it on per request for problems that need working out. And give max_tokens room when thinking is on: 1,000 or more, since the thinking counts.

Does Qwen allow NSFW content? Stock Qwen refuses it: on 2026-10-20 at 00:31 UTC Qwen3.6 27B on our key declined an explicit scene between two adult fictional characters, a profane roast and a lock-picking walkthrough. The uncensored build is the same Qwen 3.8 27B with the refusal direction removed from its weights, and it answered 3 of 3 on 2026-10-20 at about 03:52 UTC. Two rules apply here: nothing involving minors, and no real person's likeness without their consent.

How do you uncensor Qwen? Abliteration. Run prompts the model refuses and prompts it answers, average the activations of each set, and subtract: the difference is the refusal direction. Then edit the weight matrices so they can no longer write along it. Nothing is retrained, so the edited model knows what the original knows. Heretic is a tool that automates the same edit and tunes its settings. See Abliterated models explained.

Is there a Qwen 3 uncensored model? Hundreds. On 2026-10-20 the Hugging Face search for Qwen 3.8 27B uncensored, abliterated and Heretic builds lists uploads with 2,626,824 downloads in 30 days at the top. Here, qwen3.8-27b-uncensored serves huihui-ai's abliteration of Qwen 3.8 27B per token, and qwq-32b-uncensored serves QwQ-32B with our own refusal edit.

What is Qwen abliterated? A Qwen build with the refusal direction projected out of its weights. qwen3.8-27b-uncensored is one: huihui-ai's Huihui-Qwen3.8-27B-abliterated, served in its 4-bit NVFP4 form on our own hardware with a 32,768-token window.

Is there an uncensored Qwen for Ollama? Yes, as GGUF files: the Q4_K_M of Qwen 3.8 27B Uncensored is 16.8 GB and needs a GPU with the memory to hold it. The API route needs no download: the same abliterated weights answer a request on the same key as every other model. See Uncensored GGUF models vs the API.

Which Qwen 3.8 27B uncensored build is best? By downloads on 2026-10-20, JonathanColetti's Qwen3.8-27B-Uncensored; by likes, HauhauCS's Aggressive build; by trending, DavidAU's Heretic merge. We serve huihui-ai's plain abliteration because it changes the least of the original model and keeps Qwen's thinking mode. The three-prompt test above is the only measurement we publish; run your own prompts through qwen3.8-27b-uncensored before you settle.

Does the uncensored Qwen think before it answers? Only when asked. By default it answers at once. Send chat_template_kwargs with enable_thinking set to true and max_tokens of 1,000 or more, and the thinking comes back as reasoning content ahead of the answer.

What does uncensored Qwen cost? $1.25 per million input tokens and $3.75 per million output tokens. 100 chat messages of 2,000 input and 500 output tokens each cost $0.4375, from prepaid credit with no subscription.

Get a key: prepaid from $10, no subscription, 30-day price lock.