Back to Models
Google Deep MindConditional Open

TranslateGemma 12B

TranslateGemma 12B, developed by Google DeepMind, is a foundation model specialized for translation. It is equipped with a long context window of 128K and supports advanced translation tasks.

Parameters

13.0B

Context Window

128K

License

Gemma License

Release Date

2026-01-15

API Pricing

API pricing for this model is not yet available

Strengths

  • High specialization in translation
  • 128K broad context window
  • Developed by Google DeepMind

Weaknesses

  • Unknown adaptability to general tasks
  • Load from ~25GB model size
  • Subject to usage terms restrictions

Use Cases

  • High-precision translation of long documents
  • Localization of multilingual content
  • Building specialized translation pipelines

Deep Analysis

Parameters

12B

Mid-size TranslateGemma variant

Performance

Outperforms Gemma 3 27B baseline

Uses less than half the parameters to exceed 27B baseline quality

Languages

55 core / ~500 extended

Comprehensive language coverage

Benchmark

WMT24++ MetricX

Tested on 55-language WMT24++ dataset

Release Date

January 15, 2026

Part of TranslateGemma family launch

Multimodal

Yes

Image text translation via Vistra benchmark

Training

SFT + RL

Two-stage: supervised fine-tuning + reinforcement learning with ensemble reward models

Strengths

  • Outperforms Gemma 3 27B baseline on translation with less than half the parameters
  • Best balance of quality and efficiency in the TranslateGemma family
  • 55 core languages with multimodal image translation
  • Distilled from Gemini models for high fidelity
  • Open weights, available on Kaggle and Hugging Face
  • Suitable for server-side deployment with good throughput

Weaknesses

  • Lower quality than 27B variant for the most demanding translation tasks
  • Not a general-purpose model (translation-focused)
  • Extended language pairs lack confirmed metrics
  • Larger than 4B, not suitable for mobile
  • Limited community benchmarking outside Google's evaluations

Competitor Comparison

ModelArenaSWEGPQAPrice
TranslateGemma 12BN/AN/AN/AFree (open weights)
TranslateGemma 27BN/AN/AN/AFree (open weights)
Gemma 3 27B (baseline)~1430N/A~70%Free (open weights)
NLLB-200 (54.5B)N/AN/AN/AFree (open weights)
GPT-5 (general)~1480~80%~90%$5/$20 per 1M

TranslateGemma 12B is the mid-size variant that outperforms the Gemma 3 27B baseline on translation quality using less than half the parameters. It represents the best efficiency breakthrough in the TranslateGemma family, achieving high-fidelity translation across 55 languages.

Analysis generated: 2026-05-24