The Upstage consortium has been officially selected as one of Korea’s Top 3 teams advancing the country’s sovereign foundation AI initiative, a major milestone in Korea’s national AI strategy. Notably, Flitto plays a core role within the consortium by leading large-scale dataset construction, a responsibility that directly shapes model performance, reliability, and real-world usability.
The selection was announced by Korea’s Ministry of Science and ICT following the first-stage evaluation of the Sovereign AI Foundation Model Project. The evaluation applied strict criteria focused on model originality, end-to-end training capability, and long-term ecosystem impact, ultimately advancing only three teams: LG AI Research, SK Telecom, and Upstage.
In this process, Naver Cloud and NC AI were eliminated, with evaluators concluding that Naver Cloud did not fully meet the project’s core requirement for sovereign originality, a decisive factor in the final outcome.
Evaluation Criteria and the Meaning of a Sovereign Foundation Model
The first-stage evaluation was conducted using a weighted, multi-dimensional framework:
- Benchmark Performance (40%) – Measuring model capability across NIA benchmarks, global common benchmarks, and global comparative benchmarks.
- Expert Evaluation (35%) – Conducted by external AI experts reviewing architecture design, training processes, and development roadmaps.
- User Evaluation (25%) – Assessing real-world usability, inference efficiency, and deployment feasibility.
A critical component across all categories was sovereign independence, the ability to design, train, and operate a foundation model from scratch, rather than relying on fine-tuned foreign models or reused pretrained weights.
Upstage achieved top scores in global comparative benchmarks, demonstrating frontier-level performance despite relatively efficient parameter scale, which evaluators highlighted as a key competitive strength
What Is a Sovereign Foundation Model?
A sovereign foundation model refers to a large-scale AI model that is:
- Designed independently, from architecture to training strategy
- Pretrained on self-secured and self-processed datasets
- Free from restrictive foreign licenses
- Fully controllable and improvable at the national level
This policy stance reflects the Korean government’s broader commitment to technological sovereignty, national security, and long-term AI self-reliance, positioning sovereign foundation models as critical national infrastructure rather than incremental extensions of existing foreign systems.
Upstage Consortium Structure
The Upstage consortium brings together industry, academia, and research partners, each contributing specialized expertise:
- Upstage: Overall model architecture design and foundation model development
- lablup: AI infrastructure software company specializing in GPU virtualization and orchestration
- Nota AI: Compressed AI solutions and software optimization platform business with a focus on the B2B and B2G markets
- Flitto: Dataset preprocessing, quality evaluation, and large-scale language data construction
- KAIST & Sogang University faculty: Research collaboration, talent development, and international academic publication
- Industry partners across healthcare, manufacturing, legal, public sector, education, and finance to support real-world deployment
This structure reflects a full-cycle approach, from data and infrastructure to deployment and ecosystem expansion.
Flitto’s Role: Dataset Construction as a Strategic Advantage
Within the Upstage consortium, Flitto is fully responsible for dataset construction, including:
- Large-scale multilingual data preprocessing
- Quality evaluation and filtering for AI training suitability
- Alignment of training data with real-world usage scenarios
This role is strategically significant because training data quality ultimately defines model performance, especially for sovereign AI systems intended for public, industrial, and national-level use.
With more than a decade of experience building AI language datasets for global enterprises, Flitto’s participation ensures that Upstage’s foundation model is trained on high-fidelity, ethically sourced, and performance-optimized data, reinforcing both technical competitiveness and long-term sustainability.
In a project where independence, reliability, and real-world impact are paramount, Flitto’s data expertise functions as a critical pillar supporting Korea’s sovereign AI ambitions.