How to fine-tune FunctionGemma and run it locally

Abstraction: Fine-tuning Google's 270M FunctionGemma locally with Unsloth

Key points:

• FunctionGemma is a 270M-parameter Gemma variant designed specifically for task-specific tool calling, not general dialogue
• Unsloth reduces fine-tuning VRAM and time via optimized LoRA workflows; runs on NVIDIA/AMD/Intel GPUs but not Apple Silicon
• Two GGUF conversion paths: Unsloth's native conversion (Q8_0/F16/BF16) or merge-then-convert locally
• After conversion, import the .gguf file into LM Studio via lms import <path>; serve via GUI or lms server start
• Before fine-tuning, model fails to call tools; after even 10 minutes of training it correctly invokes a Wikipedia search tool
• LM Studio exposes an OpenAI-compatible local API for programmatic use of the fine-tuned model

Connections: Google · Lm Studio · Unsloth · Fine Tuning · Tool Calling · Quantization

Source: https://lmstudio.ai/blog/functiongemma-unsloth