hanemay/smartsightCoach

VISIONConcurrent Unit Cost:1Model Size:5.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 30, 2026License:cc0-1.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The hanemay/smartsightCoach model is a 5.1 billion parameter LoRA fine-tuned version of Google's Gemma-4-E2B-it-qat-q4_0-unquantized base model, developed by Niclas Bade. Optimized for on-device fitness and nutrition coaching, this vision-capable model excels at tasks like body fat estimation from photos, meal photo analysis, and nutrition label reading. It is specifically designed for mobile inference, providing an AI coach that runs entirely on-device without server round-trips.

Loading preview...

Overview

hanemay/smartsightCoach is a series of LoRA fine-tuned models, developed by Niclas Bade, for the SmartSight AI coach application. This 5.1 billion parameter model is based on Google's gemma-4-E2B-it-qat-q4_0-unquantized checkpoint, specifically chosen for its 4-bit quantization-aware training, making it suitable for efficient on-device inference on Android/iOS. The model is continuously refined through iterative updates, with each version (e.g., v85, v69) representing a promoted champion that has been rigorously evaluated against previous iterations and a held-out suite of data.

Key Capabilities

  • Vision-enabled Coaching: Integrates text coaching with advanced vision capabilities for analyzing body fat from photographs, meal photos, and nutrition labels.
  • Body Fat Estimation: Achieves improved accuracy in body fat percentage estimation, particularly at the higher end, with a mean absolute error (MAE) of 1.69 and a slope of 1.043 on real photographs (v85).
  • Conversational AI: Handles coach questions and conversational follow-ups effectively, maintaining performance while improving vision tasks.
  • On-Device Performance: Designed for .litertlm conversion, enabling the AI coach to run entirely on mobile devices without requiring server round-trips or user accounts.
  • Continuous Improvement: Each version addresses specific weaknesses, such as dietary rule compliance, conversational consistency, and accurate interpretation of historical data.

Good For

  • On-device AI applications: Ideal for mobile applications requiring local, privacy-preserving AI capabilities.
  • Fitness and Nutrition Coaching: Specifically tailored for use cases involving personalized fitness advice, dietary analysis, and progress tracking through visual input.
  • Vision-Text Multimodal Tasks: Suitable for scenarios where both natural language understanding and image analysis are crucial for providing comprehensive user support.

Finetunes
1 models