VIDRAFT’s VKUE engine runs a 34.7B-parameter sparse Mixture-of-Experts reasoning model (about 3B active parameters per token) from a single quantized weight file across hardware tiers, from 18,057 tokens/second on a datacenter B200 GPU down to roughly 17 tokens/second on CPU-only hardware. By keeping only active experts in fast memory and storing inactive experts in system RAM, the same model file runs on an 8GB laptop or a GPU-less server without re-quantization.
