mewse/gemma-4-E2B-home-assistant

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:5.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 14, 2026License:gemmaArchitecture:Transformer0.0K Featherless Exclusive Cold

mewse/gemma-4-E2B-home-assistant is a 5.1 billion parameter Gemma 4 model fine-tuned by mewse for natural language smart home control. It is optimized to generate natural language responses and emit tool calls for Home Assistant, building on the google/gemma-4-E2B-it-qat-q4_0-unquantized base. This model excels at translating user requests into actionable smart home commands, making it ideal for home automation applications.

Loading preview...

Overview

mewse/gemma-4-E2B-home-assistant is a 5.1 billion parameter Gemma 4 model, specifically a fine-tune of Google's gemma-4-E2B-it-qat-q4_0-unquantized checkpoint. Its primary purpose is to enable natural language control of smart homes, similar to the acon96/home-llm project. The model is designed to take a system prompt describing available devices and services, then respond in natural language while also emitting the necessary tool calls to execute the user's request.

This model is a complete, merged checkpoint, meaning the LoRA adapter has been integrated back into the base weights. While the base model is multimodal (text + vision + audio), only the text/language-model projections were fine-tuned, preserving the original vision and audio towers.

Key Capabilities

  • Natural Language Smart Home Control: Translates user commands into specific actions for Home Assistant.
  • Tool Call Generation: Emits the required tool calls alongside natural language responses to control smart devices.
  • Quantization-Aware Training: Inherits QAT calibration from its base, making it suitable for clean quantization to Q4_0 for llama.cpp / Ollama deployment.
  • Efficient Fine-tuning: Trained using 16-bit LoRA on the acon96/Home-Assistant-Requests-V2 dataset, focusing on learning to call tools rather than hallucinating responses.

Good For

  • Developers building natural language interfaces for Home Assistant.
  • Integrating advanced voice or text commands into smart home setups.
  • Applications requiring a model optimized for tool use in a home automation context.

Finetunes
1 models