TranslateGemma
#2 in KI-Översettengoogle · siet 2026-01-15 · 10× · tolest 30. Juni 2026
63
Momentum
TranslateGemma is a family of open-weight machine translation models by Google, built on the Gemma 3 architecture and available in three parameter sizes (4B, 12B, 27B). The models are specifically fine-tuned for translation tasks across 55 languages using a two-stage training process (Supervised Fine-Tuning + Reinforcement Learning) that distills knowledge from larger Gemini models. TranslateGemma inherits Gemma 3's multimodal capabilities, including translation of text within images. The 12B model outperforms the much larger Gemma 3 27B baseline on the WMT24++ benchmark.
Momentum-Verloop
04.04.03.07.
Features
| License | Gemma Terms of Use (Google); permits many commercial uses with Responsible AI restrictions; consent required on Hugging Face |
| Platform | Hugging Face, Kaggle, Google Vertex AI (Model Garden), Ollama (local execution); runs locally on mobile devices (4B), consumer laptops (12B), and cloud/H100 GPU (27B) |
| Price | Open-weight models available for free (download via Hugging Face / Kaggle); Vertex AI deployment billed per GCP pricing model |
| Release Date | January 15, 2026 |