README.md

metadata

title: VKIE — VIDRAFT Kernel Inference Engine
emoji: 🚄
colorFrom: purple
colorTo: indigo
sdk: static
pinned: true
license: apache-2.0
short_description: VKIE (비키) — VKAE accelerates, VKUE saves, VKIE maximizes

🚄 VKIE (비키) — VIDRAFT Kernel Inference Engine

The unified control tower for VIDRAFT's inference line, across three axes:

• 🏎️ VKAE — GPU acceleration (the sports car): ~9× faster single-stream on a B200.
• 🚗 VKUE — GPU savings (the compact car): 34.7B on a FREE CPU, no GPU.
• 🚄 VKIE — accel + savings = max serving (the train): 18,057 tok/s on one B200.

Every number is measured; every demo is live (they run on their real hardware, embedded or linked here). The "after" figures come from VIDRAFT's optimized serving — results are public, engine internals are proprietary.

Links: VKAE · VKUE · GPU-vs-CPU · CPU-only · FREE CPU · Image (T4) · Model · Blog.