This article presents Qwen3.8-27B-Humanlike-Chat, a model fine-tuned to reduce AI-like conversational habits and mimic realistic human-to-human chats. Trained on a dataset of real conversations, it aims to make responses more natural and less assistant-like.
I made this because I was getting genuinely annoyed at trying to have a normal conversation with LLMs. Even with prompting and various tricks, most models I've tried still have this "AI assistant" vibe to them that is so familiar: too helpful, polished, verbose, using words we never use in conversation, etc. I wanted a model that could just talk to me like a person, so I did the slightly unreasonable thing and put together a dataset and trained one. The dataset used for training is 125,217 obfuscated human-to-human messages across 1396 chat conversations. The goal wasn't to make Qwen smarter or improve benchmark scores. I was trying to change its conversational habits, to make it stop turning every reply into an explanation, agreeing with everything, and writing stuff just to keep the conversation "going". I trained a rank-256 LoRA on top of huihui-ai/Huihui-Qwen3.8-27B-abliterated. The released version is checkpoint 863. In my testing it feels noticeably less like an assistant, particularly in casual conversations, even without a system prompt. Replies are generally shorter, less polished, and, well, more human. There may be a tradeoff. An earlier iteration scored five percentage points lower than its Huihui parent on IFEval, an instruction-following benchmark. I haven't rerun that benchmark on this version of the checkpoint, and I haven't tested coding performance, so I don't want to pretend that number applies here. I've added a side-by-side comparison using the same system prompt, user messages, and generation settings for both models. Each model continued its own conversation branch, with reasoning effort set to 'xhigh'. Merged GGUFs and the standalone F32 LoRA adapter are in the model repo: https://huggingface.co/LessThanThreeAI/Qwen3.8-27B-Humanlike-Chat-GGUF Space where you can have a demo chat with different system prompts and reasoning modes: https://huggingface.co/spaces/LessThanThreeAI/Qwen3.8-27B-Humanlike-Chat There's also a free, rate-limited OpenAI-compatible endpoint: Base URL: https://api.lessthanthreeai.com/v1 Model:qwen3.8-27b-humanlike-chat
This article describes a drop-in chat template fix for Qwen AI models that improves accuracy, reduces token usage, and speeds up responses for knowledge work and coding tasks.
Qwen releases Qwen-AgentWorld-35B-A3B, a native language world model that simulates agentic environments across seven domains via long chain-of-thought reasoning. The model is trained with a three-stage pipeline and supports MCP, Search, Terminal, SWE, Android, Web, and OS interactions.