translategemma-4b-it-f32-GGUF

translategemma-4b-it-f32-GGUF

TranslateGemma-4b-it from Google is a lightweight 4 billion-parameter instruction-tuned translation model from the TranslateGemma family, built on Gemma 3 architecture and trained on TPUv4p/v5p/v5e hardware using supervised fine-tuning (SFT) on human/Gemini-synthesized parallel corpora across 500+ language pairs followed by RL optimization with MetricX-QE and AutoMQM reward models, enabling high-fidelity text-to-text and image-to-text translation across 55 languages (including Spanish, French, Chinese, Hindi, Arabic, and low-resource ones) with a 2K-128K token context window and 896x896 image resolution support (256 tokens/image). Optimized for edge/mobile deployment on laptops/desktops/cloud with half the parameters of baselines yet rivaling 12B models (MetricX↓=5.32, COMET↑=81.6 on WMT24++ across 55 langs), it uses a highly opinionated JSON chat template requiring source_lang_code/target_lang_code and processes inputs like {"type": "text/image_url", "text"/"url": content} for outputs in target languages, democratizing state-of-the-art translation for developers via Transformers/vLLM. Available in 4B (mobile/edge), 12B (laptops), and 27B (cloud/H100) sizes under open license, it prioritizes efficiency, throughput, and low latency without quality compromise for multilingual/multimodal communication.

translategemma-4b-it [GGUF]

Quants Usage

(sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants)

Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):