Gemma-4-31B-Humanlike-Chat Released with Natural Voice and Improved Tool-Calling
Developer LessThanThreeAI released Gemma-4-31B-Humanlike-Chat, a community fine-tune of Google's Gemma 4 31B model designed to communicate with a natural human writing style. It significantly improves tool-calling accuracy and instruction-following benchmarks compared to the official base model. This release bridges the gap between rigid AI assistants and authentic human-like conversational partners while addressing common challenges in local LLM tool use. It offers a powerful open-source alternative for developers seeking models that can naturally text and reliably handle function calling. Trained using a two-step recipe involving Supervised Fine-Tuning (SFT) and on-policy distillation with 13,740 rows of data, the model achieved a 26% success rate in blind judge tests against real human texts. It is available in various quantization formats like GGUF, Safetensors, and GPTQ-Int4 on Hugging Face.
## BACKGROUND
Fine-tuning allows developers to adapt pre-trained large language models for specific conversational tones or tasks using targeted datasets. Benchmarks like the Berkeley Function Calling Leaderboard (BFCL) evaluate how accurately models can invoke external tools and APIs, while GGUF is a portable file format optimized for running local LLM inference efficiently.