Index-Translate vs TranslateGemma: which one to pick?
Index-Translate covers 150 languages under Apache 2.0 and follows glossary and tone instructions. TranslateGemma covers 55 under the Gemma Terms and has far more downloads. We found no independent comparison.
Verdict. Choose by license and by how you give instructions. Index-Translate (2B, 9B and a 35B-A3B preview, 150 languages) is plain Apache 2.0, was built to follow glossary, format and tone instructions, and ships with a 262,144-token context. TranslateGemma (4B, 12B and 27B, 55 evaluated languages) comes from Google under the Gemma Terms of Use, with a strict language-code template, no documented glossary support and a 2K-token input context, but with far more adoption. In the Index team's own tables the 9B is ahead of TranslateGemma-12B on eight of nine columns, by far on instruction following, which TranslateGemma's documented template does not offer. TranslateGemma-12B is slightly ahead on one reference-based instruction benchmark score and has a lower off-target rate on the report's low-resource FLORES test, 1.7% against 4.0%. Google's own evaluation uses other metrics, and we found no independent comparison. Translate a sample of your own text with both.
Specifications
| Specifications | Index-Translate | TranslateGemma |
|---|---|---|
| Developer | Bilibili | Google DeepMind |
| Released | September 2026 | January 15, 2026 |
| Total parameters | 36B (Index-Translate-35B-A3B (preview, MoE)) | 27.4B (TranslateGemma 27B) |
| Active parameters | 3B | All |
| Architecture | Dense | Dense |
| Context window | 256K tokens | 2K tokens |
| Input | Text, Image | Text, Image |
| License | Apache 2.0 (Permissive) | Gemma Terms (Conditional) |
| Memory for the smallest variant (4-bit) | 1.44 GB (Index-Translate-2B (GGUF Q4_K_M)) | 2.97 GB (TranslateGemma 4B) |
Performance claims come from the developer and have not been independently verified.
Pick Index-Translate if
- You need plain Apache 2.0, with no use policy to pass on and no gated download.
- You want to give a glossary, keep JSON, Markdown or placeholders unchanged, or set a tone and genre.
- You translate languages beyond TranslateGemma's 55 evaluated ones, or inputs longer than its 2K tokens.
- You want a 2B model with a 1.31 GB GGUF file for a laptop.
Pick TranslateGemma if
- You want Google's model and the larger community: Hugging Face showed about 1.1 million downloads of the 4B model and about 960,000 of the 12B since release, against about 92,000 for the Index-Translate 9B GGUF.
- Your pipeline passes ISO language codes and plain text and needs no instructions, so the strict template is a feature.
- Off-target output matters to you, because TranslateGemma-12B had a lower rate on the Index report's low-resource FLORES test.
- You accept the Gemma Terms of Use and want a model whose quality Google also checked with human evaluation on WMT25 in 10 language pairs.