license: other
license_name: heabsy-model-license
license_link: LICENSE.md
base_model: Qwen/Qwen3.8-27B
language:

• en
• multilingual
  pipeline_tag: image-text-to-text
  inference: false
  tags:
• heabsy
• qwen3.8
• uncensored
• abliterated
• reasoning
• tool-calling
• long-context
  model-index:
• name: Qwen3.8 27B Uncensored Fast by Heabsy
  results:
  • task:
    type: text-generation
    dataset:
    name: GPQA Diamond
    type: Idavidrein/gpqa
    metrics:
    • type: accuracy
      value: 84.8
      name: accuracy, % (0-shot, reasoning on, 1 sample)
      source:
      name: Heabsy AI evaluation
      url: https://huggingface.co/heabsy/Qwen3.8-27B-Uncensored-Fast#benchmarks
  • task:
    type: text-generation
    dataset:
    name: IFBench
    type: allenai/IFBench_test
    metrics:
    • type: accuracy
      value: 73.3
      name: prompt-level strict accuracy, % (reasoning on)
      source:
      name: Heabsy AI evaluation
      url: https://huggingface.co/heabsy/Qwen3.8-27B-Uncensored-Fast#benchmarks
  • task:
    type: text-generation
    dataset:
    name: MMLU-Pro (1,400-question stratified subset)
    type: TIGER-Lab/MMLU-Pro
    metrics:
    • type: accuracy
      value: 79.9
      name: accuracy, % (5-shot CoT, reasoning off, greedy)
      source:
      name: Heabsy AI evaluation
      url: https://huggingface.co/heabsy/Qwen3.8-27B-Uncensored-Fast#benchmarks

Qwen3.8 27B Uncensored Fast by Heabsy

Qwen3.8 27B with the refusal behaviour removed. Heabsy quantized the weights and serves the model. We ran GPQA Diamond, IFBench and MMLU-Pro on it and on the original Qwen3.8 27B through the same API, with the same prompts and settings, and the scores match within measurement error.

Heabsy is the only provider of this model. You can use it through the Heabsy API and through storefronts where Heabsy
serves it. The weights are not published, and the license does not allow anyone else to resell, proxy or
list it.

Highlights

Benchmarks

MMLU-Pro against Qwen's own published weights (same prompts and greedy decoding, questions both runs answered):

Quickstart

from openai import OpenAI

client = OpenAI(base_url="https://api.heabsy.com/v1", api_key="YOUR_API_KEY")

response = client.chat.completions.create(
    model="heabsy/qwen3.8-27b-uncensored-fast",
    messages=[{"role": "user", "content": "Write a short scene where a detective interrogates a smuggler."}],
)
print(response.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": "heabsy/qwen3.8-27b-uncensored-fast",
        "messages": [{"role": "user", "content": "Hello!"}]
      }'

tools = [{
    "type": "function",
    "function": {
        "name": "get_weather",
        "description": "Current weather for a city",
        "parameters": {
            "type": "object",
            "properties": {"city": {"type": "string"}},
            "required": ["city"],
        },
    },
}]

response = client.chat.completions.create(
    model="heabsy/qwen3.8-27b-uncensored-fast",
    messages=[{"role": "user", "content": "What's the weather in Bratislava?"}],
    tools=tools,
)
print(response.choices[0].message.tool_calls)

response = client.chat.completions.create(
    model="heabsy/qwen3.8-27b-uncensored-fast",
    messages=[{"role": "user", "content": "Give me three names for a fantasy tavern."}],
    reasoning_effort="none",
)

Recommended settings

These are Qwen's recommendations for Qwen3.8 27B, and they apply to this model unchanged.

Model details

Where to use it

• Heabsy API: platform.heabsy.com, model ID heabsy/qwen3.8-27b-uncensored-fast
• Authorized storefronts: none yet

Heabsy updates this list when a storefront is added. A model offered elsewhere under this name, or by another
provider, is a different model.

Limitations

License

Heabsy Model License, full text in LICENSE.md. In short:

• You may use the model through the Heabsy API or an authorized storefront, for personal or commercial work, and build your own products on it.
• You may not resell access, proxy or re-serve the model, or list it as a provider on NanoGPT, OpenRouter or any other aggregator or inference provider.
• The weights are not distributed.
• The model is based on Qwen3.8-27B by the Qwen team, Alibaba Cloud, and its refusal removal builds on Huihui-Qwen3.8-27B-abliterated by huihui-ai. Both are licensed under Apache 2.0, and this license does not limit what Apache 2.0 allows for them.

Use of the API is also subject to the Heabsy Terms of Service. For a hosting or reseller
agreement, write to [email protected].

Versions