VIDRAFT · Inference Engine

VKUEVIDRAFT Kernel Ubiquitous Engine

No GPU? Runs anyway. Everywhere. The ubiquity engine that serves frontier-class models on minimal hardware.

VKAE — Speed ↗ VKUE — Ubiquity · Anywhere ▶ Live: GPU vs CPU ↗ ▶ Live: CPU-only ↗

18,057 tok/s

Ourbox-35B on a single B200 — VKAE speed ceiling. Measured, aggregate.

20 tok/s

Same model on an 8 GB laptop — VKUE ubiquity floor. Measured, 3.7× a dense 32B.

744B → 1 GPU

Even GLM-5.2-class giants on a single H100 (normally 6–8).

90.9%

Darwin-398B-JGOS top GPQA Diamond (greedy, no test-time engine).

VKUE Efficiency Leaderboard

How little hardware a given quality needs — 27B and up. GPQA values are each model card’s posted figure (method noted).

Model	GPQA Diamond	Size (total / active)	Standard HW	VKUE Min HW
GLM-5.2 SYMBOLIC Z.ai · 744B MoE · frontier	frontier official figure — citation pending	744B / 40B	H100 ×6–8 (INT4)	H100 ×1 ≈ 1/6–1/8 capacity verified · speed pending
Darwin-398B-JGOS Darwin V9 · 397B MoE	90.9 greedy · no engine	397B / 17B	pure GPU ×3–4 (INT4)	1 GPU + RAM available now
Darwin-28B-REASON Darwin · DELPHI engine	89.39 maj · DELPHI engine	26.9B	1 GPU	24GB ×1 CPU-capable
Darwin-28B-Opus Darwin · 28B	88.89 posted figure	~28B	1 GPU	24GB ×1 CPU-capable
★ Darwin-36B-Opus base ← Qwen3.6-35B-A3B	88.4 GPQA Diamond	36B / 3B	1 GPU (FP8)	consumer 24GB ×1 CPU-capable
Darwin-27B-Opus Qwen3.5 · 27.8B	86.9 posted figure	27.8B	1 GPU	24GB ×1 CPU-capable
★ Ourbox-35B-JGOS MEASURED Darwin-36B derivative · GGUF ↗	86.4 / 70.7 maj@8 / greedy (measured)	35B / 3B	1 GPU (FP8)	A10G 126 · CPU 17 tok/s live demo ↗ live measured · also 8GB laptop 20 tok/s (3.7× dense 32B)
Darwin-31B-Opus base ← Gemma4-31B · Korean #1	85.9 posted figure	32.7B	1 GPU	consumer 24GB ×1 CPU-capable
Qwen3.6-35B-A3B base model · Darwin-36B parent	— official figure — pending	35B / 3B	1 GPU (FP8)	consumer 24GB ×1
Gemma4-31B base model · Darwin-31B parent	— official figure — pending	31B	1 GPU	consumer 24GB ×1

Method labels · greedy = single pass, maj@8 = majority of 8 samples, DELPHI = our test-time engine. Different conditions — not directly comparable.

VKUE Min HW · INT4, held fully in VRAM or served with little / no GPU. CPU-capable = runs without a GPU (slower) thanks to ~3B active params.

Throughput (tok/s) · Ourbox-35B measured: 18,057 aggregate on B200 (VKAE ↗) / 126 on one A10G & 17 on a GPU-less CPU box (live demo ↗) / 20 on an 8GB laptop. All single-stream, generate-format–gated.

Base models · Qwen3.6-35B / Gemma4-31B are our bases; their GPQA figures will be cited from official sources.

Engine Siblings

VKAE — Speed ↗

How fast. VIDRAFT-optimized serving pushes throughput to the ceiling — Ourbox-35B measured at 18,057 tok/s aggregate on a single B200. Open the Speed leaderboard ↗

VKUE — Ubiquity ↗

How low the bar. Same 34.7B model, minimal hardware — live measured 126 tok/s on one A10G vs 17 tok/s on a GPU-less CPU box (identical weights, only the hardware differs), and 20 tok/s on an 8 GB laptop. If you have a card it’s fast; if you don’t, it still runs. Try the live GPU vs CPU demo ↗

What’s open. Quantized models, benchmark figures, and open-tool run instructions are public — anyone can reproduce the baseline. VKUE’s engine internals are proprietary; this leaderboard reports only the resulting hardware, never the how.