A Korean logistics firm's in-house AI milestone signals growing enterprise demand for CPU-based on-device inference.
TL;DR: Smart logistics company Ourbox has ranked second on the Korean government's K-AI Leaderboard with its self-developed large language model, Ourbox-31B-JGOS. As part of its broader AI transformation strategy, Ourbox is partnering with Korean AI startup VIDRAFT to deploy VIDRAFT's CPU-based on-device AI platform, POCKET, across its fulfillment centers nationwide. The collaboration, reported by ZDNet Korea on July 27, 2026, aims to bring AI capabilities to warehouse floors without relying on expensive GPU servers or external cloud infrastructure.
VIDRAFT, the Korean AI startup behind the CPU-native on-device platform POCKET, is stepping into the logistics sector through a new partnership with Ourbox — a tech-driven fulfillment company that just claimed the No. 2 spot on South Korea's government-operated K-AI Leaderboard. The collaboration was announced July 25, 2026, and covered by ZDNet Korea two days later.
VIDRAFT's role in this story centers on POCKET, its CPU-based on-device AI platform designed to run AI models in standard computing environments — removing the dependency on high-cost GPU servers and external cloud services. Ourbox, which operates end-to-end e-commerce logistics including receiving, warehousing, inventory management, picking, packing, shipping, and last-mile delivery, is now exploring how POCKET can be deployed across its warehouse network.
According to ZDNet Korea, Ourbox is evaluating a portable mini-PC form factor for POCKET installations at key operational stations — including inspection, picking, and packing zones — within its fulfillment centers. The intended outcome is a lower-cost, more flexible AI deployment model that can be scaled rapidly and tailored to the specific workload profile of each warehouse location.
The immediate backdrop is Ourbox's own AI achievement: its proprietary Korean-language large language model, Ourbox-31B-JGOS, earned second place in the 30B-and-above category on the K-AI Leaderboard, which is administered by South Korea's Ministry of Science and ICT and the National Information Society Agency (NIA). The model scored a composite average of 0.612, just 0.009 points behind the first-place model at 0.621. Standout scores included 0.971 on the CLIcK benchmark — which evaluates understanding of Korean cultural and linguistic context — along with 0.742 on Com2-main (Ko), 0.717 on KMMLU-Pro, and 0.570 on MuSR (Ko).
Ourbox CEO Park Cheol-su stated that the leaderboard result represents more than a benchmark achievement — it reflects the company's capacity to independently develop AI technology and build a foundation for solving real operational challenges in logistics. He also emphasized that the company's AI ambitions are focused on technology that workers and customers can tangibly experience, not on showcasing AI for its own sake.
The Ourbox–VIDRAFT partnership is a meaningful signal for the broader Korean enterprise AI landscape. Most industrial AI deployments to date have concentrated compute-heavy workloads in the cloud or on-premise GPU clusters — options that carry substantial infrastructure costs and are difficult to scale quickly across distributed physical sites like warehouse networks.
POCKET's value proposition addresses this friction directly. By enabling capable AI inference on commodity CPU hardware, VIDRAFT is making it possible for a logistics operator like Ourbox — which runs fulfillment centers across South Korea — to deploy AI at the individual workstation level. That means warehouse staff could, in principle, access AI-assisted guidance for tasks such as inventory queries, order verification, or operational Q&A without any cloud round-trip or specialized hardware.
For VIDRAFT, the Ourbox partnership validates POCKET as a platform suited to demanding, high-frequency industrial environments. Logistics is a sector where latency, data privacy, and cost per query all matter at scale — making it an important proving ground for CPU-based on-device AI.
Ourbox plans to apply its AI capabilities first in areas with high near-term feasibility: product and order information retrieval, inventory and outbound status Q&A, worker task assistance, client inquiry handling, standard operating procedure analysis, and logistics document automation. Longer-term applications include order volume forecasting, inventory anomaly detection, task prioritization recommendations, and decision support for warehouse operations management.
Founded in 2017, Ourbox has built its logistics platform around a proprietary warehouse management system called MATE and has expanded its service verticals to include food and cold-chain, health supplements, household goods, cosmetics, and fashion.
Q: What is VIDRAFT's POCKET platform and how is it being used in logistics?
A: POCKET is a CPU-based on-device AI platform developed by Korean AI startup VIDRAFT that allows AI models to run on standard computing hardware without GPU servers or cloud infrastructure. Ourbox is evaluating POCKET for deployment on portable mini-PCs at fulfillment center workstations for tasks like inventory queries and worker assistance.
Q: What did Ourbox's Ourbox-31B-JGOS model achieve on the K-AI Leaderboard?
A: Ourbox-31B-JGOS ranked second in the 30B-and-above large model category on the K-AI Leaderboard, operated by South Korea's Ministry of Science and ICT and NIA, with a composite average score of 0.612 — just 0.009 points behind the top-ranked model.
Q: Why is a CPU-based AI platform significant for warehouse environments?
A: GPU servers and cloud-based AI solutions carry high infrastructure costs and can be difficult to deploy quickly across many distributed physical locations. A CPU-based platform like POCKET allows logistics companies to scale AI to individual workstations faster and at lower cost, which is especially relevant for operators running large warehouse networks.
Source: ZDNet Korea (2026-07-27) — original article