EryriLabs/gemma-4-e4b-cymraeg-v4

VISIONConcurrent Unit Cost:1Model Size:7.9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 18, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

EryriLabs/gemma-4-e4b-cymraeg-v4 is a 7.9 billion parameter conversational Welsh fine-tune of Google's Gemma 4 E4B model, designed for efficient deployment on edge devices. This model excels at natural, everyday Welsh conversation, including correct mutations and idiomatic vocabulary, while retaining strong English capabilities. It supports general-purpose Welsh and English conversation, drafting, and question answering, making it suitable for applications requiring bilingual interaction without internet connectivity.

Loading preview...

Overview

EryriLabs/gemma-4-e4b-cymraeg-v4 is a conversational Welsh fine-tune of Google's Gemma 4 E4B, an "effective 4B" model optimized for edge devices like phones, laptops, and Raspberry Pi. This model is designed to run efficiently with 5 to 9 GB of memory, enabling offline Welsh language processing.

Key Capabilities

  • Bilingual Fluency: Speaks natural, everyday Welsh with correct mutations and idiomatic vocabulary, while maintaining strong English capabilities. It responds in the language of the user's input.
  • Edge Deployment: Built for environments with no internet, with quantised GGUF files available for llama.cpp and LM Studio.
  • Robust Conversation: Achieves a multi-turn conversation score of 3.99/5, demonstrating strong coherence and memory over extended interactions.
  • Commercial Use Permitted: The base model and all training data sources are commercially usable under their respective open licenses.

Training & Performance

The model underwent a two-stage LoRA training process: continued pretraining on 43.8 million tokens of open Welsh data, followed by supervised fine-tuning on 19,443 Welsh instruction rows. Evaluation shows 100% replies in the requested language for Welsh prompts, with strong grammar (3.76/5) and task completion (2.70/5) in single-turn scenarios. It significantly outperforms the stock gemma-4-E4B-it model in Welsh language tasks and conversational ability.

Good For

  • General-purpose Welsh and English conversation.
  • Drafting and question answering in both languages.
  • Applications requiring offline, on-device bilingual AI capabilities.
  • Use cases where natural Welsh idiom and correct grammar are crucial.

Finetunes
2 models