RadimSvetlak/Gemma3-270M-Pralinka-V1.1

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.3BQuant:BF16Context Size:32kPublished:Sep 10, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

RadimSvetlak/Gemma3-270M-Pralinka-V1.1 is a 0.3 billion parameter Gemma 3-based model, specifically fine-tuned for Czech language. It functions as a character model, embodying "Pralinka," a haflinger mare, to answer questions about horses and stable life. This model is optimized for running on CPUs or low-end GPUs, providing an accessible solution for domain-specific, character-driven interactions in Czech.

Loading preview...

Model Overview

RadimSvetlak/Gemma3-270M-Pralinka-V1.1 is a specialized 270 million parameter language model built upon the Gemma 3 architecture. It has undergone continued pretraining on Czech language data and subsequent supervised fine-tuning to adopt the persona of "Pralinka," a haflinger mare. This model is designed to answer questions related to horses, stable operations, training, and trail rides, specifically in Czech.

Key Capabilities

  • Czech-language character model: Responds in the persona of Pralinka, a mare with opinions.
  • Domain-specific knowledge: Excels at answering questions about horses and stable life.
  • Lightweight: At approximately 540 MB (bf16), it runs efficiently on CPUs or low-end GPUs.
  • Greedy decoding recommended: Optimized for reliable, terse answers, with a median answer length of around 65 characters.

Training and Performance

The model was developed through a three-stage full fine-tuning process, starting with google/gemma-3-270m-it, followed by continued pretraining on Czech, and finally supervised fine-tuning on a Pralinka dialogue dataset. An in-house benchmark of 73 questions across six categories demonstrated 100% accuracy within its domain, with no errors or empty outputs.

Limitations and Use Cases

  • Character, not a knowledge base: Answers are in role and not verified facts; not suitable for critical decisions (e.g., veterinary advice).
  • Strictly domain-specific: Produces nonsense outside the horse domain.
  • Small model constraints: Limited context, multi-step reasoning, and arithmetic capabilities.
  • Czech only: Not trained or evaluated for other languages.
  • Quantization sensitivity: Performance degrades noticeably below Q8_0 due to its small size.

Ideal Use Cases

  • Interactive educational tools for children or stable visitors about horses.
  • Engaging, character-driven chatbots for horse-related topics in Czech.
  • Applications requiring a lightweight, domain-specific Czech language model for defined interactions.

Finetunes
1 models