Back to Models
Google Deep MindConditional Open

TranslateGemma 4B

TranslateGemma 4B is a translation-specialized foundation model developed by Google DeepMind. It has a parameter scale of approximately 5 billion and supports a long context window of 128K.

Parameters

5.0B

Context Window

128K

License

Gemma License

Release Date

2026-01-15

API Pricing

API pricing for this model is not yet available

Strengths

  • High performance specialized in translation
  • 128K broad context capacity
  • Efficient 5B parameter scale

Weaknesses

  • General tasks beyond translation
  • 10GB model file size
  • Subject to limited usage terms

Use Cases

  • Multilingual translation of long documents
  • Translation processing of specialized text
  • Building efficient translation pipelines

Deep Analysis

Parameters

4B

Smallest TranslateGemma variant, suitable for mobile inference

Languages

55 core / ~500 extended

Core: high/mid/low-resource. Extended: nearly 500 additional pairs

Performance

Rivals Gemma 3 12B baseline

4B model matches the translation quality of the 12B base model

Base Model

Gemma 3

Distilled from Gemini models via SFT + RL

Release Date

January 15, 2026

Announced by Google

Multimodal

Yes

Retains Gemma 3 multimodal capabilities for image text translation

License

Gemma license

Available on Kaggle and Hugging Face

Training

SFT + RL

Supervised fine-tuning on parallel data + reinforcement learning with MetricX-QE and AutoMQM reward models

Strengths

  • Rivals Gemma 3 12B baseline translation quality at 1/3 the parameters
  • Supports 55 core languages and ~500 extended language pairs
  • Multimodal: can translate text within images (Vistra benchmark)
  • Runs on mobile hardware - ideal for on-device translation
  • Distilled from Gemini models for high quality
  • Open weights, available on Kaggle and Hugging Face

Weaknesses

  • Lower quality than 12B and 27B TranslateGemma variants
  • Limited to translation tasks (not a general-purpose model)
  • Extended language pairs (~500) lack confirmed evaluation metrics
  • May struggle with very complex or domain-specific translations
  • Not as widely benchmarked as general-purpose models

Competitor Comparison

ModelArenaSWEGPQAPrice
TranslateGemma 4BN/AN/AN/AFree (open weights)
TranslateGemma 12BN/AN/AN/AFree (open weights)
TranslateGemma 27BN/AN/AN/AFree (open weights)
NLLB-200 (3.3B)N/AN/AN/AFree (open weights)
Gemma 3 27B (baseline)~1430N/A~70%Free (open weights)

TranslateGemma 4B is the smallest variant in Google's open translation model family, rivaling the Gemma 3 12B baseline's translation quality with only 4B parameters. Supports 55 core languages with multimodal image text translation. Ideal for mobile and edge deployment.

Analysis generated: 2026-05-24