GigaAM-v3

Sber's 220M to 240M Russian speech recognizer, with punctuated output in its e2e versions. MIT license.

In short. GigaAM-v3 turns Russian speech into text, with versions that also add punctuation and write numbers the way people read them. It is small, with files of about 0.45 GB, and the MIT license allows commercial use.

Summary

GigaAM-v3 is the third generation of Sber's GigaAM speech models for Russian, released on 19 November 2025. It is a Conformer of 220M to 240M parameters pre-trained with a HuBERT-CTC objective on 700,000 hours of Russian speech, and the GitHub table lists 4,000 hours for the speech recognition fine-tuning. The repository stores five checkpoints as branches: ssl, ctc, rnnt, e2e_ctc and e2e_rnnt. The two e2e versions write punctuated, normalized text directly, and the rest output plain words. Sber says v3 added call-center, background-music, natural-speech and atypical-speech data, and cut errors by about 30% on those new domains while matching v2 on public benchmarks.

On the card's five test groups, the RNN-T model has an average word error rate of 8.4% and the CTC model 9.2%, against 19.4% for T-One with a language model and 25.1% for Whisper. The gaps are widest on disordered speech (19.2% against 59.3% for Whisper) and call-center audio (9.5% against 23.9%). For the e2e models, Sber ran a side-by-side judged by Gemini 2.5 Pro against Whisper large v3, and says they win by 70 to 30 on average. These are Sber's own tests, and the table covers Russian only.

The weights are MIT, and each checkpoint is about 0.45 GB. Loading needs trust_remote_code, and the card pins Transformers 4.57.1 and PyTorch 2.8. The GitHub package notes that transcribe handles audio of up to 25 seconds, and its long-form mode relies on pyannote.audio and a Hugging Face token. In the 30 days to 11 November 2026, Hugging Face counted about 245,000 downloads. The same team's GigaAM Multilingual, listed separately, adds Kazakh, Kyrgyz and Uzbek and drops the punctuation-writing e2e option.

Good for

  • Russian speech recognition with punctuation and number normalization built in, using the e2e versions.
  • Call-center and voice-message audio, where Sber's v3 data added coverage.
  • Running a small model, since each checkpoint is under 0.5 GB.
  • Commercial use under MIT, and fine-tuning from the ssl encoder.

Watch out for

  • It is a Russian model, so it is not a general multilingual recognizer.
  • All scores are Sber's own, and the Whisper comparison in the e2e test was judged by an LLM.
  • Audio over 25 seconds needs the long-form setup with pyannote.audio and a Hugging Face token.
  • The card pins older library versions (Transformers 4.57.1, PyTorch 2.8), so use a separate environment.

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-v3. 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 0.44 GB 0.506 GB Laptop or phone
e2e RNN-T (punctuation and normalization) 240M 0.45 GB 0.517 GB Laptop or phone
RNN-T 240M 0.45 GB 0.517 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.

More about MIT

See also

Frequently asked questions

What is GigaAM-v3?
GigaAM-v3 turns Russian speech into text, with versions that also add punctuation and write numbers the way people read them. It is small, with files of about 0.45 GB, and the MIT license allows commercial use.
Can I use GigaAM-v3 commercially?
Yes. GigaAM-v3 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-v3 open source?
GigaAM-v3 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-v3 on my own computer?
Yes. The smallest variant, CTC, needs about 0.506 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was GigaAM-v3 released?
GigaAM-v3 was released by Sber (GigaChat) on November 19, 2025.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
GigaAM Multilingual Sber (GigaChat) 220M – 600M – MIT (Permissive) 1.01 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
Parakeet TDT 0.6B v3 NVIDIA 627M – CC BY 4.0 (Permissive) 2.89 GB Laptop or phone