Qwen3.8-27B-Humanlike-Chat
An open-source model fine-tuned from Qwen3 27B using LoRA, with human-like conversational style as its core characteristic. The model supports tool calling, role-playing, and instruction following. Its design objective is to enable large language models to communicate in a manner resembling natural human expression rather than the formulaic tone typical of AI assistants.
Timeline (last 90 days)
The original model received 700+ upvotes and 248 comments on Reddit, with a download count of 44,000.
The original model had three major shortcomings: it could only produce a few words at a time, had a single default persona that could not be changed via prompting, and completely lacked tool-calling ability.
The original version used standard supervised fine-tuning (SFT), with data drawn from 139,845 real and synthetic messages across 1,396 conversations.
The model trained a rank-256 LoRA adapter on top of the base model huihui-ai/Huihui-Qwen3.8-27B-abliterated.
The final released version is checkpoint 863.
After fine-tuning, the model's responses are generally shorter, less polished in wording, and overall closer to real human expression.
An early iteration scored five percentage points lower than the Huihui base model on the IFEval instruction-following benchmark.
A Reddit developer (LessThanThreeAI) trained a model specifically designed to imitate real human conversation, called Qwen3.8-27B-Humanlike-Chat.
This type of style fine-tuned model is better suited for companionship chat and role-playing scenarios, rather than production environments that require rigorous task execution.
The author releases resources in four forms: a merged GGUF, a standalone F32 LoRA adapter, an online Demo Space, and a free OpenAI-compatible API endpoint.
5 more timeline events
All Facts (14)
The original model had three major shortcomings: it could only produce a few words at a time, had a single default persona that could not be changed via prompting, and completely lacked tool-calling ability.
50%UnverifiedThe original version used standard supervised fine-tuning (SFT), with data drawn from 139,845 real and synthetic messages across 1,396 conversations.
50%UnverifiedThe original model received 700+ upvotes and 248 comments on Reddit, with a download count of 44,000.
50%UnverifiedThe model trained a rank-256 LoRA adapter on top of the base model huihui-ai/Huihui-Qwen3.8-27B-abliterated.
50%UnverifiedThe final released version is checkpoint 863.
50%UnverifiedAfter fine-tuning, the model's responses are generally shorter, less polished in wording, and overall closer to real human expression.
50%UnverifiedAn early iteration scored five percentage points lower than the Huihui base model on the IFEval instruction-following benchmark.
50%UnverifiedThis type of style fine-tuned model is better suited for companionship chat and role-playing scenarios, rather than production environments that require rigorous task execution.
50%UnverifiedIn the side-by-side comparison test, both models use the same system prompt, user messages, and generation parameters, with the reasoning intensity set to 'xhigh'.
50%UnverifiedThe author has not yet re-run the IFEval benchmark on the currently released checkpoint, nor has coding performance been tested.
50%UnverifiedThe training objective of the model was not to improve intelligence or benchmark scores, but to change conversational habits so that the model speaks more like a human.
50%UnverifiedThe author collected 125,217 desensitized real human conversation messages spanning 1,396 chat records for training.
50%UnverifiedThe author releases resources in four forms: a merged GGUF, a standalone F32 LoRA adapter, an online Demo Space, and a free OpenAI-compatible API endpoint.
50%UnverifiedA Reddit developer (LessThanThreeAI) trained a model specifically designed to imitate real human conversation, called Qwen3.8-27B-Humanlike-Chat.
50%