Best open-source translation models in 2026

Open translation has moved from general chat models to specialist ones. Four families of text translators are now available, from Google, Tencent, Bilibili and Preferred Networks, and they differ most in language coverage, in what they accept as instructions and in their licenses. The list then covers a model that translates text found inside images and three speech models that translate audio. Every entry says what it is built for and what its license asks of you.

  • List updated November 11, 2026
  1. Hy-MT2 Most downloaded

    Tencent's Apache 2.0 family in 1.8B, 7B and 30B-A3B sizes for 33 languages. The 7B GGUF had about 1.16 million downloads in the 30 days to 11 November 2026, and a 1.25-bit build of the 1.8B model is only 440 MB.

    • Tencent
    • 30.1B (3B active)
    • 256K tokens
    • Apache 2.0 (Permissive)
    • from 1.41 GB
  2. Index-Translate

    Bilibili's Apache 2.0 family in 2B, 9B and 35B-A3B (preview) sizes, with 150 languages, image translation and glossary, format and tone controls. In its own report the 9B leads Hy-MT2-7B and TranslateGemma-12B on eight of nine columns, which is the team's own test.

    • Bilibili
    • 36B (3B active)
    • 256K tokens
    • Apache 2.0 (Permissive)
    • from 1.44 GB
  3. TranslateGemma

    Google's Gemma 3-based 4B, 12B and 27B models for 55 languages, which also translate text in images. Hugging Face showed about 1.1 million downloads of the 4B model since its January 2026 release. The Gemma Terms of Use apply, and the model card gives a total input context of 2K tokens.

    • Google DeepMind
    • 27.4B
    • 2K tokens
    • Gemma Terms (Conditional)
    • from 2.97 GB
  4. PLaMo 3 Translate Newest

    The newest entry, released on 5 November 2026. Preferred Networks' 0.8B to 31B models cover 53 languages, are tuned for Japanese and English, and have GGUF files from 0.57 GB. Commercial use needs registration and stops at 1 billion yen of revenue without a separate agreement.

    • Preferred Networks
    • 31.8B
    • 32K tokens
    • PLaMo Community (Conditional)
    • from 0.614 GB
  5. HunyuanOCR

    Tencent's 1.1B OCR model that also translates the text it finds in menus, posters and screenshots, which a text-only translation model cannot read. Its license does not cover the EU, UK or South Korea.

    • Tencent
    • 1.12B
    • 128K tokens
    • Tencent Hunyuan (Conditional)
    • from 0.773 GB
  6. Granite Speech 5.0

    IBM's Apache 2.0 2B model that transcribes six languages and translates speech to and from English in them, plus English to Italian and Mandarin. The BF16 weights are about 5 GB.

    • IBM
    • 2.31B
    • Apache 2.0 (Permissive)
    • from 1.09 GB
  7. Nemotron 3.5 Transcribe

    NVIDIA's speech model, about 34B parameters in total, that translates speech to or from English in 31 of its 40 languages and can attribute speakers in 13. It is 22 GB in NVFP4 and comes under the OpenMDW-1.1 license.

    • NVIDIA
    • 34B (4.3B active)
    • 64K tokens
    • OpenMDW-1.1 (Permissive)
    • from 25 GB
  8. LFM3-Audio-2B

    Liquid AI's 2B on-device voice model: speech translation across 18 languages together with transcription and speech synthesis, without sending audio to the cloud. Its license allows commercial use only below US$10 million of annual revenue.

    • Liquid AI
    • 2B
    • 32K tokens
    • LFM Open 1.0 (Conditional)
    • from 4.6 GB

How we chose

  • Open weights you can download today from the developer's official repository.
  • The text translators come first, ordered by how simple their license is and then by adoption. Image and speech models follow.
  • Quality scores come from the developers' own tests and use different benchmarks, so we do not rank by them. We found no independent comparison of these text translators.
  • We read each license in full, because four of these models limit commercial use by revenue, region or a use policy.