Model reference · open weights
Borealis is an open-weight audio or speech model from Vikhrmodels. Borealis (BF16) weighs 2.3 GB; the smallest configuration that runs it is RTX 3060 12 GB.
For self-hosting, compare this build's requirements below with the GPU you plan to rent.
Summary of the Vikhrmodels/Borealis model card, 2026-10-01
What it is
| Released by | Vikhrmodels |
|---|---|
| Released | 2025-09-11 |
| VRAM | 2.3 GB for the weights |
What it runs on
| Card | One stream | Memory |
|---|---|---|
| RTX 3060 12 GB | fits | 11.6 GB |
| RTX 4060 Ti 16 GB | fits | 15.4 GB |
| RTX 3090 24 GB | fits | 23.4 GB |
| RTX 4090 24 GB | fits | 23.4 GB |
| RTX 5090 32 GB | fits | 31.0 GB |
| L40S 48 GB | fits | 44.0 GB |
| A100 80 GB | fits | 78.2 GB |
| H100 80 GB | fits | 78.1 GB |
| RTX PRO 6000 Blackwell 96 GB | fits | 93.8 GB |
| DGX Spark (GB10) 128 GB unified | fits | 107 GB |
| H200 141 GB | fits | 138 GB |
| B200 180 GB | fits | 176 GB |
From the model card
Borealis - это наша первая audio llm c ASR для русского языка. В этом репозитории представлен чекпоинт, который видел примерно 7000 часов аудио на русском языке. Важным отличием от других моделей является то, что тут есть поддержка пунктуации в распознанных аудио. Архитектура во многом вдохновлена Voxtral, но отличается в некоторых моментах.
from transformers import AutoModelForCausalLM, AutoTokenizer, AutoFeatureExtractor
import torch
import librosa
model = AutoModelForCausalLM.from_pretrained("Vikhrmodels/Borealis", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("Vikhrmodels/Borealis")
extractor = AutoFeatureExtractor.from_pretrained("Vikhrmodels/Borealis")
generation_params = {
"max_new_tokens": 350,
"do_sample": True,
"top_p": 0.9,
"top_k": 50,
"temperature": 0.2,
}
model.eval()
model.to("cuda")
waveform, sr = librosa.load("path/to/your/audio.wav", sr=16_000)
proc = extractor(
waveform,
sampling_rate=sr,
padding="max_length",
max_length=480_000,
return_attention_mask=True,
return_tensors="pt",
)
mel = proc.input_features.squeeze(0).to("cuda")
att_mask = proc.attention_mask.squeeze(0).to("cuda")
with torch.inference_mode():
transcript = model.generate(mel=mel, att_mask=att_mask, **generation_params)
print(transcript)
Ниже представлены замеры Borealis на фоне остальных открытых моделей, который поддерживают русский язык. Бенчмарк мы скоро выложим в открытый доступ
| Модель | Средний WER | RuLS | CV 22.0 | Books | Speak | Sova |
|---|---|---|---|---|---|---|
| Borealis | 6.33% | 6.39% | 2.67% | 5.28% | 1.95% | 15.37% |
| GigaAM-ASR-V2-RNNT | 5.85% | 5.24% | 2.85% | 8.06% | 3.08% | 10.01% |
| openai/whisper-large-v3 | 10.74% | 11.62% | 7.51% | 12.19% | 2.74% | 19.65% |
| bond005/whisper-podlodka-turbo | 9.38% | 11.91% | 6.36% | 8.96% | 3.14% | 16.55% |
| openai/whisper-large-v3-turbo | 11.30% | 11.88% | 8.17% | 13.29% | 2.80% | 20.37% |
| bond005/whisper-large-v3-ru-podlodka | 10.76% | 10.24% | 7.80% | 10.31% | 3.23% | 22.21% |
| nvidia/canary-1b-v2 | 13.52% | 20.16% | 9.12% | 11.45% | 4.97% | 21.89% |
| VOSK-model-ru-0.42 | 11.30% | 12.06% | 11.87% | 10.80% | 2.61% | 19.15% |
| GigaAM-ASR-V2-CTC | 6.45% | 5.26% | 3.42% | 7.72% | 3.01% | 12.86% |
Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.