NAKSTStudio/chess-gemma-commentary

TEXT GENERATIONConcurrent Unit Cost:1Model Size:0.3BQuant:BF16Context Size:32kPublished:Nov 3, 2025License:gemmaArchitecture:Transformer0.0K Featherless Exclusive Cold

NAKST Studio's chess-gemma-commentary is a 0.3 billion parameter Gemma 3 270M model fine-tuned for generating multilingual chess move commentary, ELO predictions, and move classifications. Optimized for mobile and offline inference, it supports 14 languages and provides detailed analysis of chess positions and moves. This model excels at integrating into applications requiring on-device, privacy-focused chess analysis.

Loading preview...

What is NAKSTStudio/chess-gemma-commentary?

This model is a Gemma 3 270M variant, fine-tuned by NAKST Studio using LoRA on over 25,000 chess positions with expert commentary. It specializes in generating detailed chess move commentary, predicting ELO ratings (1000-2800), and classifying moves (e.g., Best Move, Blunder). A key differentiator is its multilingual support, offering commentary in 14 languages including English, Hindi, Spanish, and Mandarin Chinese.

Key Capabilities

  • Multilingual Chess Commentary: Provides analysis in 14 languages.
  • ELO Prediction: Estimates player skill ratings.
  • Move Classification: Labels moves with tags like 'Best Move', 'Mistake', 'Blunder'.
  • Mobile & Offline Ready: Designed for lightweight inference on Android devices via flutter_gemma or Ollama, requiring no internet connection.
  • Conversational Input: Utilizes a structured system and user message format for precise input of FEN, MoveSAN, and other chess data.

Why is this model different?

Unlike general-purpose LLMs, chess-gemma-commentary is hyper-specialized for chess analysis, offering specific outputs like ELO predictions and move classifications. Its small size (0.3B parameters) and support for quantized .task files make it ideal for on-device, privacy-focused applications where larger models are impractical. The model's strict input/output format ensures consistent, structured responses, making it highly reliable for automated chess analysis within applications.