Ourbox-31B-JGOS Hits #2 on Korea's K-AI Leaderboard — with VIDRAFT's POCKET Bringing CPU On-Device Inference to Warehouse Floors

#kaileaderboard #koreanllm #ondeviceai #logisticsai

Ourbox-31B-JGOS Hits #2 on Korea's K-AI Leaderboard — with VIDRAFT's POCKET Bringing CPU On-Device Inference to Warehouse Floors

TL;DR: Smart logistics company Ourbox has self-developed a 31B-parameter Korean LLM called Ourbox-31B-JGOS, which ranked #2 overall on the government-operated K-AI Leaderboard (30B+ category) as of July 25, 2026. In parallel, Ourbox is collaborating with Korean Pre-AGI AI startup VIDRAFT to deploy VIDRAFT's POCKET CPU-native on-device AI platform inside physical fulfillment centers — eliminating GPU server dependencies on the warehouse floor.

What it is

Ourbox-31B-JGOS is a large language model developed entirely in-house by Ourbox, a Korean e-commerce fulfillment company founded in 2017. The model targets Korean language understanding and reasoning at a competitive level, benchmarked on the K-AI Leaderboard — an official evaluation framework operated by South Korea's Ministry of Science and ICT (MSIT) and the National Information Society Agency (NIA).

The 30B+ model category on K-AI Leaderboard is designed for large-scale models, and Ourbox placed #2 overall in that tier. The company intends to combine this general Korean language capability with its proprietary logistics system ("MATE") and operational data to build a logistics-specialized AI — handling tasks like inventory QA, order processing, worker assistance, and document automation.

VIDRAFT's POCKET is a separate but complementary piece of the stack: a CPU-based on-device AI platform designed to run AI models in standard CPU environments, without requiring expensive GPU servers or external cloud connectivity. Ourbox is evaluating deploying POCKET in the form of portable mini-PCs installed at key warehouse workstations — inspection, picking, and packing stations.

How it works

At a conceptual level, the Ourbox stack is a two-layer approach:

Cloud/server-side reasoning layer: Ourbox-31B-JGOS handles heavy Korean language tasks — understanding complex queries, multi-step reasoning, domain document parsing — trained on and evaluated against Korean-specific benchmarks.
Edge inference layer (VIDRAFT POCKET): POCKET abstracts away the GPU dependency by enabling model inference on commodity CPU hardware. This is aimed at scenarios where installing rack-mounted GPU servers in a fulfillment center is operationally impractical or cost-prohibitive.

The goal of combining both is a tiered deployment architecture: high-complexity tasks route to the larger model, while localized, low-latency interactions at individual workstations are served by POCKET-powered mini-PCs without cloud round-trips.

From an ML-engineering perspective, POCKET's value proposition is hardware-agnostic inference — bringing AI capability to environments with standard x86 or ARM compute rather than CUDA-capable hardware. No further architectural details of POCKET are publicly disclosed at this time.

Benchmarks & results

All figures below are sourced directly from Ourbox's public announcement (as of July 23, 2026) and the K-AI Leaderboard:

Benchmark	Score
K-AI Leaderboard Overall (30B+ tier)	0.612 (2nd place)
1st place model (for reference)	0.621
Gap to #1	0.009
CLIcK (Korean cultural & linguistic context)	0.971
Com2-main (Ko)	0.742
KMMLU-Pro	0.717
MuSR (Ko)	0.570

The CLIcK score of 0.971 is particularly notable — this sub-benchmark specifically evaluates a model's grasp of Korean cultural context and linguistic nuance, which is a known hard problem for models primarily trained on English-dominant corpora. The KMMLU-Pro score tests expert-level multi-domain knowledge in Korean, and MuSR (Ko) evaluates multi-step reasoning in Korean.

How to try it

Ourbox-31B-JGOS and VIDRAFT POCKET are not currently available as public open-source releases, Hugging Face model cards, or developer-accessible APIs based on information available at the time of this article. Ourbox has indicated the model will be deployed internally within its own logistics operations, with no public access channel announced.

If you want to follow developments:

Monitor the K-AI Leaderboard (operated by NIA) for updated rankings and any newly released model cards.
Watch VIDRAFT's official channels for any future public announcements about POCKET's developer access or SDK availability.

FAQ

Q: Is Ourbox-31B-JGOS based on an existing open-weight base model (e.g., Qwen, LLaMA)?
A: The source article does not disclose the base architecture or training lineage of Ourbox-31B-JGOS. No such details have been made public at this time.

Q: What makes VIDRAFT's POCKET different from standard quantized inference libraries like llama.cpp or ONNX Runtime?
A: The source article describes POCKET as a platform designed to run AI models on standard CPU environments without GPU servers or external cloud dependencies, but does not publicly detail its underlying inference stack or how it compares to existing open-source CPU inference tools. Further technical comparison is not possible without additional public disclosure from VIDRAFT.

Q: Does deploying on CPU inevitably mean worse latency for warehouse workers?
A: Latency is a real trade-off in CPU inference, but for many logistics edge tasks — item lookup, checklist QA, document parsing — the interaction cadence doesn't require sub-100ms response times. The value proposition here is cost and deployment flexibility, not raw throughput.

Originally reported by ZDNet Korea (2026-07-27) — source article.