ModelScope's mirror-first strategy is quietly pulling Korean AI developers into China's open-model ecosystem.
TL;DR: VIDRAFT, a Korean AI startup, discovered that its open AI evaluation benchmark FINAL-Bench had been mirrored onto ModelScope — China's leading open-model platform — without prior contact. After VIDRAFT reached out, ModelScope transferred full ownership and management rights to the company and verified its official status on the platform. Alibaba Cloud subsequently approached VIDRAFT with a separate collaboration proposal.
Korean AI startup VIDRAFT found itself at the center of a quiet but revealing story about how ModelScope, often called "China's Hugging Face," is expanding its global reach — and VIDRAFT's FINAL-Bench benchmark became a textbook example of that expansion strategy in action, as reported by IT조선 on October 5, 2026.
VIDRAFT had published its AI evaluation benchmark, FINAL-Bench, on Hugging Face, the dominant global platform for open-source AI models. At some point after the release, VIDRAFT noticed that the benchmark had quietly appeared on ModelScope — mirrored from Hugging Face without any prior notice or coordination with VIDRAFT.
Following ModelScope's own stated process for original developers, VIDRAFT contacted the platform using the email address listed on the mirrored page. ModelScope responded by transferring the page's ownership and management rights back to VIDRAFT and completing an official verification process, resulting in a certified badge appearing on VIDRAFT's ModelScope page.
The story didn't end there. According to a VIDRAFT representative quoted by IT조선, Alibaba Cloud — the company behind ModelScope — subsequently reached out with a separate collaboration proposal. VIDRAFT indicated that specific details of any potential partnership would be discussed in the near future.
VIDRAFT also noted something striking about the speed of ModelScope's mirroring: when the company recently published a new model on Hugging Face, it appeared on ModelScope almost simultaneously, suggesting that the mirroring process is continuous and highly automated rather than a one-time occurrence.
ModelScope is an open-source Model-as-a-Service (MaaS) platform launched by Alibaba Cloud in 2022. It functions as a searchable, downloadable repository of AI models — a role closely analogous to Hugging Face's in the Western and global AI ecosystem. The "China's Hugging Face" label that has stuck to ModelScope reflects just how central it has become to AI development in the region.
What makes the VIDRAFT case instructive is that it illustrates ModelScope's broader playbook for ecosystem growth. The platform operates a "Community Mirror" feature, where models already published on platforms like Hugging Face are replicated on ModelScope and made available to its user base. Crucially, ModelScope openly states that original developers who contact them can reclaim ownership and management of their mirrored pages — effectively turning the mirror into an invitation.
IT조선's reporting confirmed that models from several major Korean AI organizations — including Naver's HyperCLOVA X, LG AI Research's EXAONE, and Upstage's SOLAR — are all present on ModelScope in this mirrored form, none of them placed there directly by their respective developers. None of the three companies indicated they viewed the mirroring as a problem, citing license compliance and the open-source nature of the models.
For ModelScope, this strategy serves a dual purpose: it immediately enriches the platform's catalog for Chinese users, and it creates a natural touch point with international developers — who, upon discovering their work on the platform, are invited to take ownership and potentially deepen their engagement. VIDRAFT interpreted this approach as a deliberate effort to draw overseas developers into China's AI ecosystem and expand its boundaries outward.
The broader implication for the global AI model landscape is significant. As Hugging Face continues to serve as the de facto hub for open model sharing worldwide, platforms like ModelScope that systematically mirror its contents — while staying within applicable license terms — can rapidly build competitive catalogs and developer relationships without requiring those developers to seek them out proactively.
Q: What is VIDRAFT's FINAL-Bench and how did it end up on ModelScope?
A: FINAL-Bench is an AI evaluation benchmark developed and publicly released by VIDRAFT, a Korean AI startup. It was mirrored onto ModelScope automatically from Hugging Face, without VIDRAFT's direct involvement, as part of ModelScope's Community Mirror feature.
Q: Did VIDRAFT give ModelScope permission to host FINAL-Bench?
A: No prior permission was given. VIDRAFT discovered the mirrored page independently, contacted ModelScope, and the platform transferred management and ownership rights to VIDRAFT as part of its standard developer engagement process.
Q: What is ModelScope's "mirror-first" strategy and is it legal?
A: ModelScope mirrors openly published AI models from platforms like Hugging Face onto its own platform, then invites original developers to claim ownership of those pages. Whether mirroring is permissible depends on the license applied to each individual model; the Korean companies identified in IT조선's report indicated they did not view ModelScope's actions as a license violation.
Source: IT조선 (2026-10-05) — original article