VIDRAFT's POCKET Platform Powers CPU-Only On-Device AI in Korean Logistics Fulfillment Centers
TL;DR: VIDRAFT, a Korean AI startup, is deploying its POCKET on-device AI platform — a CPU-based, portable mini-PC solution — in partnership with logistics company Ourbox to run inference in warehouse environments without GPUs or cloud connectivity. The collaboration accompanies Ourbox's launch of its own 31B-parameter Korean LLM,
Ourbox-31B-JGOS, which ranked 2nd overall on the government-operated K-AI Leaderboard (30B+ category). Engineers working on edge inference, Korean NLP, or domain-adapted LLMs should pay attention.
What it is
This story covers two interlocking technical developments in the Korean AI ecosystem:
1. Ourbox-31B-JGOS — a 31-billion-parameter Korean large language model developed in-house by Ourbox (a smart logistics and fulfillment company). The model is purpose-built for the logistics domain, combining general Korean language understanding and complex reasoning with proprietary operational data from Ourbox's warehouse management system, MATE.
2. VIDRAFT's POCKET platform — an on-device AI platform from VIDRAFT (비드래프트), architected around CPU-based inference and packaged as a portable mini-PC. POCKET is being introduced into Ourbox fulfillment centers specifically for tasks at the inspection, picking, and packing stations.
The partnership between Ourbox and VIDRAFT represents a practical answer to a recurring problem in industrial AI deployment: how do you run real-time AI inference in environments where high-end GPUs are cost-prohibitive and network connectivity is unreliable?
How it works
Ourbox-31B-JGOS follows a domain-specialization approach that engineers will recognize: start with a strong general-purpose Korean language base (evidenced by top-tier scores on Korean comprehension benchmarks), then layer in domain-specific fine-tuning using logistics operational data — SKU information, order/inventory states, warehouse procedures, and exception-handling workflows. The model is integrated directly with the MATE warehouse management system, enabling it to query real-time operational context rather than relying on static knowledge.
Planned inference use cases follow a sensible difficulty progression:
- Near-term (retrieval + generation): product/order information lookup, worker task guidance, operations manual Q&A, document automation
- Medium-term (predictive): workload forecasting, inventory anomaly detection, task prioritization recommendations
VIDRAFT's POCKET takes a different architectural bet: rather than requiring a GPU or cloud API call, it runs AI inference entirely on CPU hardware in a compact, portable form factor. The key design goals are:
- Air-gap tolerance — stable operation even when warehouse network connectivity drops
- Cost reduction — eliminates the need for per-station GPU hardware or cloud API spend at scale
- Security — data stays on-premises, which matters for logistics clients with sensitive inventory or order data
- Portability — the mini-PC form factor means it can be repositioned across different station types (inspection, picking, packing) without infrastructure changes
The combination is architecturally coherent: a domain-adapted LLM handles the language understanding and reasoning, while POCKET handles the inference runtime at the edge.
Benchmarks & results
The source article reports the following public benchmark results for Ourbox-31B-JGOS on the K-AI Leaderboard, operated by South Korea's Ministry of Science and ICT (MSIT) and the National Information Society Agency (NIA):
| Category | Score |
|---|---|
| K-AI Leaderboard Overall Rank (30B+ category) | 2nd place |
| K-AI Leaderboard Composite Score | 0.612 |
| Korean Contextual Understanding (CLIcK) | 0.971 |
The model is also noted to score highly in professional knowledge and complex reasoning sub-categories, though specific sub-scores beyond CLIcK are not detailed in the source.
No public benchmark data for VIDRAFT's POCKET platform itself (e.g., tokens/sec, latency, specific CPU targets) is included in the press coverage.
How to try it
Neither Ourbox-31B-JGOS nor VIDRAFT's POCKET platform appear to have publicly announced Hugging Face model cards, GitHub repositories, or open API endpoints at the time of this reporting. Both are described as proprietary systems being deployed in production logistics environments.
- To explore the K-AI Leaderboard results and methodology, the leaderboard is operated by NIA under MSIT and is publicly accessible at the Korean government's AI evaluation portal.
- For developer access inquiries to either system, direct contact with Ourbox or VIDRAFT through their official channels would be the appropriate route.
This section will be updated if public access channels are announced.
FAQ
Q: Is POCKET essentially just a quantized model running on a CPU — what's actually novel here?
A: The press coverage doesn't disclose POCKET's internal inference stack or quantization approach. What is described is the system-level value proposition: a CPU-only, portable, air-gapped-capable inference appliance specifically targeting industrial floor environments where GPUs and cloud dependencies are operational liabilities. The novelty claimed is in the deployment architecture and cost profile, not necessarily a new inference algorithm.
Q: How does Ourbox-31B-JGOS relate to existing Korean LLMs — is it fine-tuned from a public base model?
A: The source article does not specify the base model or training lineage. It is described as "independently developed" (독자 개발), but whether that means trained from scratch or domain-adapted from an existing foundation is not disclosed.
Q: Why does Korean language benchmark performance matter for a logistics AI use case?
A: Korean logistics operations involve highly context-dependent, exception-heavy communication — irregular packing specs, per-client handling rules, real-time order amendments. Strong Korean contextual understanding (as measured by CLIcK) is a prerequisite for reliable instruction-following and document parsing in this environment. A model that struggles with Korean nuance would generate unsafe or incorrect operational guidance at the warehouse floor level.
Originally reported by 동아일보 (2026-07-27) — source article.
Top comments (0)