Google drops TranslateGemma with 12B model beating 27B baseline

TranslateGemma is Google's new family of open translation models built right on top of Gemma 3. It gives you research-grade multilingual translation on your own hardware, and it can even read and translate text inside images without a separate OCR pipeline.

The lineup comes in three sizes: 4B, 12B, and 27B parameters. The killer detail here is pure density. On the WMT24++ benchmark evaluated with MetricX, the 12B TranslateGemma model beats the baseline Gemma 3 27B model, while the 4B variant rivals the original 12B baseline.

What it does

Out of the box, TranslateGemma targets 55 core languages. That covers heavy hitters like English, Spanish, French, Chinese, and Hindi, alongside a solid spread of low-resource languages. Google says error rates dropped significantly across the board compared to raw Gemma.

Beyond those 55 languages, the team trained the models on nearly 500 additional language pairs. Google has not published formal benchmark scores for that extended set yet, but it gives researchers a massive head start for fine-tuning niche languages.

It also keeps the multimodal eyes of Gemma 3. On the Vistra image translation benchmark, the improved text engine naturally boosted image-in-text translations, even though Google did not run any special vision fine-tuning during this run.

How it works and hardware requirements

Google pulled off this efficiency by distilling knowledge from its flagship Gemini models into the compact Gemma architecture through a two-stage process.

First came supervised fine-tuning (SFT) on a blend of human translations and Gemini-generated synthetic pairs. Then came reinforcement learning (RL), guided by an ensemble of reward models including MetricX-QE and AutoMQM to smooth out phrasing and make the output sound natural.

Hardware needs are practical. The 4B model is tuned for on-device mobile inference. The 12B model runs smoothly on modern consumer laptops for local development. The flagship 27B model fits comfortably on a single H100 GPU or cloud TPU.

Why it matters

If you build apps that need offline translation or private multilingual pipelines, TranslateGemma changes the math. You no longer have to burn budget on proprietary APIs or drag a massive 70B general-purpose model just to translate clean text.

The 12B model hitting 27B quality on a regular laptop is a huge win for edge computing. While the long tail of 500 extra language pairs still needs community testing, anyone handling cross-border products or on-device document translation should download the 4B or 12B weights and start tinkering today.