Ourbox-31B-JGOS Ranks 2nd Overall on Korea's Official K-AI Leaderboard (30B+ Category) — Built with VIDRAFT's POCKET On-Device AI Platform
TL;DR: Ourbox's
Ourbox-31B-JGOSlarge language model has secured 2nd place overall in the 30B+ parameter category on the Korean government's official K-AI Leaderboard, administered by the Ministry of Science and ICT (MSIT) and the National Information Society Agency (NIA). The model was developed in collaboration with VIDRAFT's on-device AI platform POCKET. For developers working in Korean-language AI or on-device inference, this result signals a meaningful new entrant in the competitive Korean LLM ecosystem.
What it is
Ourbox-31B-JGOS is a large language model developed by Ourbox (아워박스), sitting in the 30-billion-plus parameter class. It achieved 2nd place overall in the 30B+ category on the K-AI Leaderboard, a public benchmark maintained by South Korea's Ministry of Science and ICT (과기정통부) and the National Information Society Agency (NIA).
The key collaboration detail: this result was achieved in partnership with VIDRAFT, a Korean Pre-AGI AI startup, specifically leveraging their POCKET on-device AI platform. POCKET is VIDRAFT's infrastructure for running capable AI models efficiently in on-device or edge-proximate environments — a meaningful technical constraint that shapes how the model is optimized.
Key facts from the source:
-
Model name:
Ourbox-31B-JGOS - Developing organization: Ourbox (아워박스)
- Platform partner: VIDRAFT (via the POCKET on-device AI platform)
- Benchmark: K-AI Leaderboard, 30B+ parameter division
- Result: 2nd place overall in its category
- Governing bodies: Korea's MSIT (과기정통부) and NIA
How it works
At a conceptual level, the collaboration between Ourbox and VIDRAFT reflects a broader engineering challenge in the LLM space: how do you train or fine-tune a 30B+ parameter model while keeping inference viable for on-device or resource-constrained deployment scenarios?
VIDRAFT's POCKET platform is described as an on-device AI platform, which implies a focus on:
- Efficient inference pipelines optimized for lower-power or edge hardware targets
- Model compression or quantization strategies that preserve benchmark performance while reducing memory and compute footprints
- Deployment abstractions that allow capable models to run outside of large cloud datacenter environments
The Ourbox-31B-JGOS model, developed within or alongside this platform, was evaluated on the K-AI Leaderboard — a standardized evaluation suite designed to measure Korean-language understanding and generation capability across diverse tasks.
It is worth noting that the specific internal training configurations, architectural modifications, and optimization techniques are not publicly disclosed. What is public is the benchmark outcome and the general collaborative framework.
Benchmarks & results
The source provides the following publicly reported result:
| Leaderboard | Category | Result |
|---|---|---|
| K-AI Leaderboard (NIA / MSIT) | 30B+ parameter division | 2nd place overall |
The K-AI Leaderboard is South Korea's official government-backed evaluation framework for large language models, making a top-3 finish in the largest publicly tracked parameter tier a significant signal of model quality for Korean-language tasks.
No additional numerical sub-scores (e.g., per-task accuracy figures) are reported in the available source material. Developers interested in the detailed evaluation breakdown should consult the official NIA K-AI Leaderboard directly once full results are published.
How to try it
Based on the available source material, no specific public access channels (Hugging Face repository, GitHub link, or API endpoint) for Ourbox-31B-JGOS are confirmed in the reporting.
Similarly, VIDRAFT's POCKET platform access details are not publicly specified in this article.
If you want to stay current:
- Watch VIDRAFT's official channels for POCKET platform announcements
- Monitor the NIA K-AI Leaderboard for any model cards or linked repositories associated with ranked models
- Check Ourbox's public developer presence for model releases
⚠️ Do not attempt to use unverified endpoints or third-party mirrors claiming to host this model — wait for official release announcements from Ourbox or VIDRAFT.
FAQ
Q: What is the K-AI Leaderboard, and why does a 2nd-place finish matter?
A: The K-AI Leaderboard is South Korea's official government-administered LLM benchmark, run by the National Information Society Agency (NIA) under the Ministry of Science and ICT. It evaluates models specifically on Korean-language tasks, making it the authoritative public ranking for Korean LLM capability. A 2nd-place finish in the 30B+ category — the largest tracked tier — is a strong signal for production use cases requiring Korean-language performance at scale.
Q: What exactly is VIDRAFT's POCKET platform, and is it available to developers?
A: POCKET is VIDRAFT's on-device AI platform, designed to enable capable AI model inference in edge or on-device contexts. Based on the current source, public developer access details for POCKET have not been announced. Developers interested in the platform should follow VIDRAFT's official channels for access programs or SDK releases.
Q: Is Ourbox-31B-JGOS comparable to other public 30B-class models like Qwen or Llama-class derivatives?
A: The source does not provide a cross-model comparison against specific public baselines. The K-AI Leaderboard ranking is Korean-language-specific, so direct comparison to general-purpose English-centric benchmarks would require additional evaluation data not present in this report.
Originally reported by 한국경제 (한경) (2026-07-27) — source article.
Top comments (0)