Introducing TranslateGemma: a groundbreaking collection of open translation models built on Gemma 3, now available in 4B, 12B, and 27B parameter sizes. This innovative suite enhances communication across 55 languages, regardless of device. By compressing advanced large model knowledge into high-performance, efficient open models, TranslateGemma ensures superior translation quality without compromising on efficiency. Notably, the 12B model surpasses the larger 27B baseline model in performance, achieving this milestone with fewer parameters, thereby enhancing throughput and reducing latency. Meanwhile, the 4B model matches the 12B baseline’s performance, making it ideal for mobile applications. Tested against the WMT24++ benchmark, TranslateGemma significantly decreases error rates across all language families, including high-, mid-, and low-resource languages. This model suite represents a substantial advancement for developers seeking high-fidelity translation solutions with improved efficiency. Embrace TranslateGemma for unparalleled translation quality and efficiency in a compact form.
Source link
