GigaAM Multilingual
Sber's 220M and 600M speech recognizers for Russian, Kazakh, Kyrgyz and Uzbek, pre-trained on 2M hours of 70+ languages. MIT.
In short. GigaAM Multilingual turns recorded speech into text and is built especially for Russian, Kazakh, Kyrgyz and Uzbek, where it beats larger, better-known recognizers in the developer's tests. It is small, with files of 0.9 GB and 2.3 GB, and the MIT license allows commercial use.
Summary
GigaAM Multilingual is a family of Conformer speech encoders of 220M and 600M parameters from Sber's GigaChat team. According to the card they were pre-trained with a HuBERT-style objective on 2 million hours of speech in more than 70 languages, then fine-tuned for recognition with character-wise CTC decoders on 50,000 hours. Its Hugging Face repository first appeared on 14 July 2026, although the GitHub news lists it under June 2026. The paper behind it, on underrepresented Central Asian languages, was submitted to arXiv on 11 July 2026, and the GitHub news says it was accepted at Interspeech 2026. There are four checkpoints, ssl, ctc, large_ssl and large_ctc, stored as branches of one Hugging Face repository, and the ssl encoders can be fine-tuned for a new language.
The card says the models give the best open-source quality on Russian, Kazakh, Kyrgyz and Uzbek and only moderate quality on English. Its word error rates, on Common Voice, FLEURS and internal in-the-wild sets with greedy decoding, show the large model at 3.0% on Russian FLEURS, 4.4% on Kazakh, 5.5% on Kyrgyz and 7.3% on Uzbek, against 3.1%, 32.4%, 86.3% and 105.4% for Whisper large v3, and 4.6%, 6.8%, 9.5% and 11.9% for Seamless M4T large v2. On English FLEURS it scores 9.4% against 3.9% for Whisper large v3. These are Sber's own tests, with utterances over 30 seconds and references containing digits excluded.
The weights are MIT, and the PyTorch files are 0.88 GB for the 220M model and 2.34 GB for the 600M one. Loading needs trust_remote_code, and the card recommends Transformers 5 with PyTorch 2.10. The GitHub package notes that transcribe works on audio of up to 25 seconds, and longer files need its long-form mode, which relies on pyannote.audio and a Hugging Face access token. In the 30 days to 11 November 2026, Hugging Face counted about 64,000 downloads.
Good for
- Transcribing Russian, Kazakh, Kyrgyz and Uzbek speech with a small model.
- Building a recognizer for another language by fine-tuning the self-supervised encoder.
- Running speech recognition with small files, 0.88 GB for the 220M model and 2.34 GB for the 600M one.
- Commercial use under MIT, without per-minute fees.
Watch out for
- English quality is only moderate, and the card shows Whisper large v3 well ahead on English FLEURS.
- The card reports only English, Russian, Kazakh, Kyrgyz and Uzbek, so check any other language of the 70-plus before relying on it.
- The card describes character-wise CTC output and mentions no punctuation or speaker labels, and audio over 25 seconds needs the long-form setup with pyannote.audio.
- All scores are Sber's own, and the card's test sets exclude references that contain digits.
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of GigaAM Multilingual. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| CTC (220M) | 220M | 0.88 GB | 1.01 GB | Laptop or phone |
| Large CTC (600M) | 600M | 2.34 GB | 2.69 GB | Laptop or phone |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.
See also
- Best open-source speech recognition (speech-to-text) models in 2026
- Sber's open models: a 432B chat model, speech and embeddings
Frequently asked questions
- What is GigaAM Multilingual?
- GigaAM Multilingual turns recorded speech into text and is built especially for Russian, Kazakh, Kyrgyz and Uzbek, where it beats larger, better-known recognizers in the developer's tests. It is small, with files of 0.9 GB and 2.3 GB, and the MIT license allows commercial use.
- Can I use GigaAM Multilingual commercially?
- Yes. GigaAM Multilingual is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is GigaAM Multilingual open source?
- GigaAM Multilingual is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run GigaAM Multilingual on my own computer?
- Yes. The smallest variant, CTC (220M), needs about 1.01 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was GigaAM Multilingual released?
- GigaAM Multilingual was released by Sber (GigaChat) on July 14, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- GigaAM Multilingual model card (Hugging Face) huggingface.co
- GigaAM repository and release notes (GitHub) github.com
- GigaAM Multilingual: Foundation Model for Underrepresented Languages (arXiv 2607.10371) arxiv.org
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| GigaAM-v3 Sber (GigaChat) | 220M – 240M | – | MIT (Permissive) | 0.506 GB Laptop or phone | |
| Whisper large-v3-turbo OpenAI | 810M – 1.54B | – | MIT (Permissive) | 6 GB Laptop or phone | |
| Qwen3-ASR Alibaba Qwen | 918M – 2.35B | – | Apache 2.0 (Permissive) | 2.12 GB Laptop or phone | |
| Cohere Transcribe Cohere | 2.07B | – | Apache 2.0 (Permissive) | 4.75 GB Laptop or phone |