A new fine-tuned version of Google’s Gemma 4 31B model, titled “Gemma-4-31B-Humanlike-Chat,” has been released by an independent developer. The model aims to replicate natural human texting patterns, moving away from the standard assistant-style responses typical of large language models (LLMs).

Technical Specifications and Training Methodology

According to reports on Reddit, the model is a fine-tune of the official google/gemma-4-31B-it base model and is released under the Apache-2.0 license. The developer utilized a two-step training process on a single H200 GPU, involving Supervised Fine-Tuning (SFT) and on-policy distillation.

The SFT phase utilized 13,740 rows of data, incorporating the developer’s previous “2.0” model replies, original Google instruction data, and human preference selections. The subsequent distillation phase involved 12 steps where the model generated its own responses, which were then rated by two “teachers”: one focused on human-like conversational style and another focused on maintaining the original Google base’s proficiency in instructions, tools, and code.

Performance Benchmarks

The developer provided several comparative benchmarks to demonstrate the model’s performance against the original Google base and the previous Qwen 2.0 fine-tune.

In terms of conversational “humanity,” the developer conducted a blind test where a judge was asked to distinguish between real human messages and model-generated messages. The developer claims the Gemma-4-31B-Humanlike-Chat model was identified as a human in 26% of instances, whereas the original Google Gemma 4 31B was identified as a human in 0% of the tests.

Regarding functional capabilities, the new model showed improvements in several areas over the Google base:

  • When2Call (correctly calling, asking, or refusing): Increased from 46 to 66.
  • BFCL Irrelevance (not calling tools when unsuitable): Increased from 81 to 90.
  • BFCL Simple (correct tool and arguments): Increased from 95 to 96.

When compared to the Qwen 2.0 model, the Gemma-4-31B-Humanlike-Chat demonstrated higher scores in LiveCodeBench (66 vs 51) and IFEval instruction following (90.0 vs 83.4). However, the developer noted that the model’s knowledge base remains consistent with the Google base, scoring 82.5 on the MMLU-Pro benchmark.

Use Cases and Community Reception

The model is designed to respond in short, lowercase lines without the need for specific prompting to adopt a persona. While the developer suggests applications in character-based dialogue for gaming, some users on social media have raised concerns regarding potential misuse in automated scams or dating app bots.

Initial feedback from the community has been mixed. While some users noted the model’s utility for creating consistent NPCs in game development, others reported that the conversational style can occasionally feel “flat” or “synthetically short.”

The model is currently available in various formats, including GGUF and Safetensors, via Hugging Face.