Qwen3.8 27B Uncensored (NVFP4)
Qwen3.8 27B Uncensored is Qwen3.8 27B with the refusal direction removed from its weights (abliteration after Arditi et al., published by OrcaRouter under Apache 2.0); vision input, tool calling and the thinking mode are kept. The publisher states it is meant for research into refusal mechanisms, red-teaming and safety evaluation, and that anything shown to end users needs its own moderation layer — the model complies with requests the original declines. We serve it on hardware we operate inside the EEA; prompts and completions are not stored.
model id: qwen38-unc
from openai import OpenAI client = OpenAI( api_key="hb-...", base_url="https://api.heabsy.com/v1",) r = client.chat.completions.create( model="qwen38-unc", messages=[{"role": "user", "content": "Hello"}],)print(r.choices[0].message.content)
curl https://api.heabsy.com/v1/chat/completions \ -H "Authorization: Bearer $HEABSY_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "qwen38-unc", "messages": [{"role": "user", "content": "Hello"}] }'
import OpenAI from "openai"; const client = new OpenAI({ apiKey: process.env.HEABSY_API_KEY, baseURL: "https://api.heabsy.com/v1",}); const r = await client.chat.completions.create({ model: "qwen38-unc", messages: [{ role: "user", content: "Hello" }],});console.log(r.choices[0].message.content);
Price list, updated 12 September 2026. No minimum spend, no subscription.
About the model
- Developer
- OrcaRouter
- Released
- August 2026
- Licence
- Apache-2.0
- Architecture
- dense · 27B parameters
- Based on
- Qwen3.8 27B
- Input
- text, images
- Reasoning
- switchable per request
- Languages named by the developer
- English, Chinese
Source: the developer's model card, as of 17 September 2026.
Which Qwen3.8 27B variant is this?
Qwen3.8 27B is the Qwen team's original model — the one for production work. It is not Qwen3.8-Max, a separate and far larger model that most news about “Qwen 3.8” refers to.
Qwen3.8 27B Uncensored (OrcaRouter) removes the refusal direction from the weights and keeps vision, tool calling and thinking; its publisher measured refusals on harmful prompts falling from 64–99 % to 0–6 % and scores on MMLU, GSM8K and CMMLU staying within about one point. Qwen3.8 27B Obliterated (OBLITERATUS) removes refusals by a different weight edit and reports about two points lost on MMLU. Both are meant for research and red-teaming.
Qwen3.8 27B Queen is tuned for roleplay and fiction; Qwen3.8 27B Fable is a light tune on assistant and coding-agent conversations. Several other “uncensored” builds of the same base exist on Hugging Face — the open-weights field under technical details names the exact repository we serve.
Licence: what it means for your company
Apache-2.0. A permissive open-source licence: you may use, modify and sell products built on the model, including commercially; when you redistribute the weights, the licence and copyright notices go with them.
A summary of the licence, not legal advice.
Technical details
- Context window
- 512K tokens
- Max output
- 32,768 tokens
- Throughput limit
- 32M tok/min
- Tokenizer
- Qwen
- Streaming
- yes
- Open weights
- orcarouter/Qwen3.8-27B-Uncensored
- max_tokens
- 1 … 32768
- temperature
- 0 … 2
- top_p
- 0 … 1
- frequency_penalty
- -2 … 2
- presence_penalty
- -2 … 2
- stop
- array
- seed
- integer
This model runs on GPUs we run inside the European Economic Area. Prompts and completions are not written to our logs or durable storage, nothing is used for training, and request content is processed in the EEA. It is the model behind every measurement on the inference page — 176 tok/s per stream, sub-second first token, 91–97% prefix-cache hit on agentic workloads.
Questions about this model
›Where is Qwen3.8 27B Uncensored computed?
Inside the European Economic Area, on hardware we operate. The catalog reports the processing region per model, so you can answer a supervisory authority without asking us.
›Does Qwen3.8 27B Uncensored store prompts and completions?
No. This model runs with zero data retention: request content is not written to durable storage and nothing is used for training.
›How large is the Qwen3.8 27B Uncensored context window?
512K tokens, with up to 32,768 tokens of output.
›What does Qwen3.8 27B Uncensored cost?
$0.20 per million input tokens and $1.20 per million output tokens. No minimum spend and no subscription; the same price list covers the API and the invoice.
›Is there a throughput limit on Qwen3.8 27B Uncensored?
Yes, 32M tokens per minute. Higher limits are a question of volume, not of principle.
›Which languages does Qwen3.8 27B Uncensored support?
The developer names English, Chinese. Languages not on that list may still work, but they are not claimed — test on your own material first.
›Under which licence is Qwen3.8 27B Uncensored released?
Apache-2.0. A permissive open-source licence: you may use, modify and sell products built on the model, including commercially; when you redistribute the weights, the licence and copyright notices go with them.
›Can reasoning be switched off in Qwen3.8 27B Uncensored?
Yes. Per the developer, thinking can be switched on or off for each request.
Compare with similar models
| Model | Developer | Parameters | Context | Licence | Input / 1M |
|---|---|---|---|---|---|
| Qwen3.8 27B Uncensored (NVFP4) | OrcaRouter | 27B | 512K | Apache-2.0 | $0.20 |
| Qwen3.8 27B (NVFP4) | Qwen (Alibaba) | 27B | 256K | Apache-2.0 | $0.40 |
| Qwen3 32B | Qwen (Alibaba) | 32.8B | 128K | Apache-2.0 | $0.12 |
| Qwen3.6 35B A3B | Qwen (Alibaba) | 35B / 3B | 256K | Apache-2.0 | $0.15 |
30 more models behind the same key and the same invoice.
Browse the catalog