The development of artificial intelligence in Russia is currently constrained by a lack of specialized infrastructure, specifically GPU clusters and data centers. According to the analysis presented at the RUSSOFT 2026 forum, domestic AI solutions are effective for simple tasks or highly specialized industry applications, but they struggle to compete with foreign models in general-purpose tasks due to limited computational resources.
Global development models for AI infrastructure vary. The Western approach relies on private capital and a circular exchange of resources among more than 16 major companies to fund hardware and data centers. The Eastern model, particularly in China, is driven by state support and big tech collaboration to achieve national goals. Russia faces the need to choose a development path to build its own sovereign large-scale models.
The ExVRAM Lab project is exploring methods to run large local LLMs on consumer hardware with limited VRAM. Using an NVIDIA GeForce RTX 5060 8 GB card based on the Blackwell architecture, researchers tested the Qwen3.8–27B model. Initial results showed that keeping model weights in system RAM resulted in speeds of 3.8–5.5 tokens per second due to PCIe data transfer bottlenecks.
The research highlights that the primary performance barrier for running dense models on consumer GPUs is not a lack of raw computing power, but the overhead of moving data between CPU and GPU. The project aims to optimize this by using ultra-low-bit quantization to fit more weights directly onto the GPU, thereby improving performance for local AI applications.
