Truthseeker87/solarhive-e4b-ollama

VISIONConcurrent Unit Cost:1Model Size:7.9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 16, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

Truthseeker87/solarhive-e4b-ollama is a 7.9 billion parameter Gemma 4 E4B model, LoRA fine-tuned by Truthseeker87 and merged into BF16 safetensors. Optimized for local deployment, it offers multimodal visual question answering and native function calling for community energy management. This model is designed for privacy-first edge deployment on consumer hardware, excelling at well-specified, in-scope queries related to solar production, grid status, and battery management.

Loading preview...

SolarHive E4B: Edge-Optimized Multimodal AI for Community Energy

SolarHive E4B is a 7.9 billion parameter Gemma 4 E4B model, LoRA fine-tuned by Truthseeker87 and merged into BF16 safetensors. It serves as the edge companion to the larger SolarHive 26B A4B model, specifically optimized for local deployment via Ollama on consumer hardware, emphasizing a privacy-first approach to community energy data.

Key Capabilities

  • Multimodal Visual Question Answering (VQA): Leverages the base Gemma 4 E4B vision encoder for sky analysis, panel inspection, and neighborhood assessment from images.
  • Native Function Calling: Integrates with 5 tools, including OpenWeatherMap, Open-Meteo GHI, simulated Community BMS, EIA Open Data, and NREL PVWatts v8, for real-time data retrieval.
  • Selective Tool Reasoning: Intelligently decides when to invoke tools based on query context, avoiding unnecessary calls.
  • GGUF Conversion Source: This repository provides the source safetensors for creating efficient GGUF artifacts, ideal for llama.cpp and Ollama deployment on systems with limited RAM (e.g., 16 GB CPU laptops).

Performance and Limitations

While designed for efficiency, the smaller E4B model exhibits a regression in refusal/follow-up behavior (2/3 on When2Call probes compared to 3/3 for the larger A4B family), indicating that it may auto-fill missing parameters rather than asking clarifying questions for underspecified queries. It achieves a converged loss of 0.9218 and a 9/10 score on a project-held-out 10-prompt parity check. The model requires ≥24 GB RAM for direct Transformers inference; for 16 GB systems, the companion solarhive-e4b-gguf repository is recommended.

Ideal Use Cases

  • Local, Privacy-Sensitive Energy Management: Deploying AI directly on-site for community solar projects without cloud dependency.
  • Well-Specified Query Handling: Excels at direct questions about solar production, grid rates, and battery status.
  • Research and Further Fine-tuning: Serves as a reference for extending LoRA on additional data using Unsloth FastVisionModel.

Finetunes
1 models