LessThanThree AI's Qwen3.8 Fine-Tune Texts Like a Human While Calling Tools
A 27B uncensored Qwen fine-tune texts like a real person without a system prompt, while still handling tools, code, and formal writing on demand.
Takeaways
- LessThanThree AI released Qwen3.8-27B-Humanlike-Chat 2.0, a GGUF chat model that texts like a person with no system prompt.
- Blind ishuman v2 judge mistook its replies for a real human 23.5% of the time versus 15.1% for official Qwen3.8-27B.
- IFBench rose from 37.3 to 43.7 and BFCL irrelevance jumped from 60 to 78 versus the abliterated base.
- Trained via on-policy distillation with two teachers, in roughly 6 hours on a single H100.
- Apache-2.0, fits a 24 GB card at IQ4_XS (15.1 GB), with a free hosted OpenAI-compatible endpoint available.
- Built on huihui-ai Huihui-Qwen3.8-27B-abliterated so content filtering is minimal; MMLU-Pro and coding dip slightly.
LessThanThree AI has released Qwen3.8-27B-Humanlike-Chat 2.0, an open-weights chat model whose Hugging Face counter had passed 62,000 downloads at publication. The 27-billion-parameter Qwen3.8 derivative ships as GGUF files for local runtimes such as llama.cpp. It starts from an “abliterated” checkpoint, a community-modified base with much of its refusal behavior weakened.
The model defaults to short, informal messages that resemble one-to-one texting. It can follow requests for formal writing, explanations, code, structured output, and tool calls, then resume its casual style on later turns. That behavior makes it relevant to persona-led chat, creative applications, and conversational agents that need function calling.
Performance figures come from the creator’s evaluations and should be reproduced under the intended runtime, quantization, language, and prompt format. The published results show stronger style imitation and several tool-use gains, alongside lower knowledge and coding scores.
Casual voice, put to a blind test
LessThanThree AI created an evaluation called ishuman v2 using 147 moments from one-to-one conversations. An LLM judge receives the preceding chat and two possible next messages, one written by a person and one generated by a model. The reported score measures how often the model’s reply is selected as the human message, with 50% treated as the point of indistinguishability.
| Model or configuration | Score |
|---|---|
| Abliterated base | 0.3% |
| Base with a human-texting prompt | 6.8% |
| Official Qwen3.8-27B | 15.1% |
| Humanlike Chat 2.0 | 23.5% |
The 23.5% result places version 2.0 well above the tested base configurations, although the evaluation remains narrow. An LLM served as judge, most chat samples were in Russian, and the test contains only 147 moments. Its scope ends at next-message mimicry, leaving user preference, factual accuracy, safety, and long-conversation consistency unmeasured.