VIDRAFT · Unified Inference Engine

VKIE 비키VIDRAFT Kernel Inference Engine

VKAE accelerates. VKUE saves. VKIE maximizes serving. One 34.7B model, from a datacenter GPU down to a free CPU — every number measured, every demo live.

The problem

Speed, cost, and scale pull against each other — win one, usually lose two.

Speed

Slow inference is unusable — but going fast normally means an expensive datacenter GPU.

💸

Cost

Serving a large model means a multi-GPU cluster. With GPU shortages, you often can't even buy them.

📈

Scale

Model and hardware get locked in — no single engine spans datacenter down to edge / on-prem.

🏎️

VKAE

GPU acceleration · the sports car

~9× faster
single-stream, 1×B200 (24 → 220 tok/s)
🚗

VKUE

GPU savings · the compact car

runs on FREE CPU
34.7B at ~6-7 tok/s, zero GPU
🚄

VKIE (비키)

accel + savings = max serving · the train

18,057 tok/s
1×B200 serving capacity (aggregate)

Before → After · measured

Each axis, on the same hardware, same 34.7B model — what our engine changes.

🏎️ VKAE — acceleration
single-stream · 1×B200
Before
24
After
220
≈ 9× faster
baseline serving → VIDRAFT optimized (tok/s)
🚗 VKUE — savings
same 8 GB laptop · dense → sparse
Before
5.4
After
20.0
≈ 3.7× faster
dense 32B → sparse A3B, identical hardware (tok/s)
🚄 VKIE — serving capacity
1×B200 · single → optimized concurrent
Before
24
After
18,057
≈ 750× serving capacity
naive single request → VKIE concurrent serving (aggregate tok/s)

Three axes at a glance

Same 34.7B model, same principles — three optimization targets.

🏎️ VKAE🚗 VKUE🚄 VKIE
FocusSpeedSavingsUnified · throughput
Best hardwareDatacenter GPUCPU ~ small GPUFull range
StrengthTop single-stream speedLowest cost · accessibilityMax tok/s · cost-efficiency
Measured before→after24→220 (9×)5.4→20 (3.7×)24→18,057 (750×)
In a phrasefastestcheapestmost

One model, the whole spectrum · measured

Ourbox-35B-JGOS — 34.7B total / ~3B active MoE. Same weights, only the hardware changes.

HardwareMeasured tok/sAxis
1× B200 (datacenter)18,057VKIE ceiling · aggregate (256 concurrent)
1× A10G (cloud GPU)126VKUE · single-stream
8 GB gaming laptop (RTX 5060)20.0VKUE · 3.7× a dense 32B
CPU-Upgrade (8 vCPU, no GPU)~17VKUE
FREE CPU space (2 vCPU, no GPU)~6-7VKUE floor · zero cost

Quality holds across every tier — GPQA Diamond 86.4% (Ourbox-35B, maj@8) up to 90.9% (Darwin-398B). Multimodal: Janus-Pro-1B image generation on a low-cost T4 in ~28 s/image (fp16, 2.1× our first cut).

Why it matters

Breaking the rule that "big AI needs big money."

🏛️

Sovereign AI

Public sector, defense, healthcare, finance — data that can't touch the cloud. Frontier reasoning on an air-gapped on-prem CPU.

💰

Cost collapse

From a multi-hundred-thousand-dollar GPU cluster to a ~$1,600 card — or a free CPU. Entry cost drops by orders of magnitude.

🌍

Accessibility

Individuals, startups, SMBs, public bodies — anyone. Ready for the surge in on-device and edge demand.

Live tests · try it yourself

These run on their real hardware (not this page). Click a card to open, or load one on screen below.

🟢
34.7B on one A10G
Normally an H100 job — VKUE runs it on a 24 GB A10G, live tok/s.
Open ↗
🔵
CPU only · no GPU
34.7B on 8 vCPU, zero GPU.
Open ↗
🆓
FREE CPU (~6-7 tok/s)
34.7B on HuggingFace's free CPU tier.
Open ↗
🖼️
Image gen (T4)
Autoregressive image on a low-cost GPU.
Open ↗
🏎️
VKAE — speed
Datacenter throughput leaderboard.
Open ↗
🚗
VKUE — efficiency
Minimal-hardware leaderboard.
Open ↗
🆓 Load VKUE on screen — 34.7B on a FREE CPU (no GPU)
Click "Load live" to embed the free-CPU demo here (34.7B, ~6-7 tok/s, zero GPU).
🟢 Load VKUE on screen — 34.7B on one A10G GPU
Click "Load live" to embed the A10G demo — 34.7B, which normally needs an H100 (wakes the A10G, ~1 min).

All links

🏎️ VKAE Speed 🚗 VKUE Efficiency 🟢 A10G · 34.7B 🔵 CPU-only 🆓 FREE CPU 🖼️ Image (T4) 🤗 Model 📝 Blog
Honest scope. Every figure above is our own measurement, reproducible on the linked demos. The "before" baselines use standard open tooling; the "after" numbers come from VIDRAFT's optimized serving — the resulting speed and hardware are public, the engine internals are proprietary. VKAE/VKIE serving numbers are on a single B200; VKUE numbers are on the exact consumer/CPU hardware named. No claim that a CPU beats a GPU — VKUE's point is that the model runs where a GPU isn't available.