Release of Qwen3.8-27B-Humanlike-Chat 2.0

A new iteration of the Qwen3.8-27B model, titled “Qwen3.8-27B-Humanlike-Chat 2.0,” has been released by developer LessThanThreeAI. The update aims to refine the conversational tone of the previous version while addressing technical limitations regarding instruction following and tool integration.

According to reports on Reddit, the initial version of the model—a LoRA (Low-Rank Adaptation) based on the abliterated Qwen3.8-27B—gained significant traction within the open-source community, recording over 44,000 downloads. However, users identified several functional drawbacks, including an inability to execute tool calls and a tendency to ignore specific formatting instructions.

Technical Improvements and Training Methodology

The 2.0 version utilizes on-policy distillation to improve performance. According to the developer, the training process involved two distinct “teachers” to guide the student model: one focused on maintaining a natural, human-like conversational tone, and another focused on technical accuracy for instructions, code, and tool usage. This method allows the model to adopt a human-like persona without requiring a specific system prompt to maintain that behavior.

Key functional updates reported by the developer include:

  • Enhanced Instruction Following: The model can switch between casual texting and formal formats, such as numbered lists or professional emails, upon request.
  • Tool Call Integration: Unlike the first version, 2.0 is designed to execute tool calls and request missing information rather than generating speculative data.
  • Coding and Mathematics: The model maintains performance levels comparable to the base model in these categories.

Benchmark Performance

Data provided by the developer suggests improvements in several key metrics compared to the abliterated base model. In instruction following (IFBench), the model reportedly increased from 37.3 to 43.7. Tool-related benchmarks also showed gains: the “When2Call” metric rose from 48 to 58, and the “BFCL irrelevance” metric—measuring the ability to refrain from calling unnecessary tools—increased from 60 to 78.

However, the developer noted that the model still trails the base model in certain areas, specifically in general knowledge (MMLU-Pro: 72.5 vs 78.5) and competitive coding (LiveCodeBench: 51 vs 56).

Human-Likeness Testing

To measure conversational authenticity, the developer utilized a custom benchmark referred to as “ishuman.” This test involves a blind evaluation where judges attempt to distinguish between real human messages and model-generated messages.

In this test, the Qwen3.8-27B-Humanlike-Chat 2.0 was identified as human by judges 23.5% of the time, compared to 15.1% for the official, unmodified Qwen3.8-27B model. In a separate test involving 16 multi-turn chats with invented personas, the developer claims the 2.0 model was preferred over the base model in all 16 instances.

The model is currently available in GGUF formats and as a standalone LoRA on Hugging Face.