Borealis audio LLM trained from scratch for under $2,000
A Russian-language audio language model trained on consumer hardware demonstrates that speech-capable LLMs no longer require enterprise budgets.
Borealis is an audio-capable language model for Russian trained from scratch on hardware costing roughly the price of a MacBook. Developer Ilya Wortega began the project a year ago during an internship at Vikhri and recently completed training, documentation, and vLLM integration for inference. The model represents a shift in accessibility for multimodal AI — what once required enterprise compute budgets now runs on consumer-grade GPUs.
The project targets audio LLM training rather than text-to-speech synthesis, a distinction Wortega emphasizes as simpler and more approachable for practitioners. Audio LLMs process and generate speech as part of their language understanding pipeline, while TTS systems focus narrowly on converting text to audio waveforms. The former requires less specialized acoustic modeling and can leverage existing transformer architectures with audio tokenization layers.
Full training details, architecture notes, and bilingual documentation (Russian and English) are published on HuggingFace Spaces at huggingface.co/spaces/AlexWortega/borealis-blog. The model runs on vLLM for production inference, a choice that signals practical deployment intent rather than research-only experimentation. vLLM's PagedAttention and continuous batching make it the de facto standard for serving open-weight LLMs at scale, and Wortega notes the integration work as a point of pride in the project writeup.
Audio-capable models have historically lagged behind text and vision in the open-source ecosystem, partly because training them required expensive multi-GPU clusters and specialized audio datasets. Projects like Whisper democratized speech recognition, but generative audio LLMs remained out of reach for individual researchers. Borealis demonstrates that the compute barrier has dropped enough for a single developer to train a functional model over the course of a year on hardware that costs less than a high-end laptop. The project's total compute cost stayed below $2,000, putting multimodal speech training within reach of independent researchers and small teams. Russian-language models remain underserved compared to English, making Borealis a notable contribution to non-English open-source AI infrastructure.