ModelScope's Mirror Strategy: What Korean AI Benchmark FINAL-Bench's Experience Tells Developers About Cross-Platform Model Distribution
TL;DR: ModelScope (Alibaba Cloud's open-source MaaS platform, often called "China's Hugging Face") actively mirrors publicly released models from Hugging Face — including VIDRAFT's
FINAL-BenchAI evaluation benchmark — and then hands ownership back to original developers who reach out, often followed by a formal collaboration offer. If you've published models or datasets on Hugging Face under a permissive license, there's a meaningful chance they're already on ModelScope, and you can claim them.
What it is
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 — conceptually analogous to Hugging Face's Hub — and is one of China's primary public AI model platforms.
VIDRAFT is a Korean AI startup that publishes FINAL-Bench, a publicly released AI evaluation benchmark. VIDRAFT's FINAL-Bench became a concrete case study in how ModelScope's mirroring pipeline operates in practice.
Other Korean models confirmed to appear on ModelScope via the same mirroring mechanism include:
- Naver — HyperCLOVA X SEED variants (Think-32B, Think-14B, Omni-8B, Text-Instruct-1.5B, Text-Instruct-0.5B, Vision-Instruct-3B)
- LG AI Research — EXAONE 3.5, EXAONE Deep, EXAONE 4.0, K-EXAONE (including K-EXAONE 236B-A23B and EXAONE Deep 2.4B)
- Upstage — SOLAR-10.7B, Solar Pro, and Solar Open family models
Importantly, none of these three organizations directly registered their models on ModelScope themselves — all confirmed the listings were created by ModelScope's mirroring process.
How it works
ModelScope operates a "Community Mirror" page system with a clearly defined handoff process:
Automatic mirroring: ModelScope crawls or monitors public model repositories (primarily Hugging Face) and mirrors models and datasets that carry licenses permitting redistribution. It creates a named page on ModelScope — often matching the original organization/repo naming — and makes weights directly downloadable.
Continuous sync: Mirroring appears to be ongoing rather than one-time. VIDRAFT noted that when they published a new model on Hugging Face, the same model appeared on ModelScope with almost no delay — suggesting an automated pipeline rather than a manual curation process.
Ownership transfer offer: ModelScope's Community Mirror pages include explicit instructions stating that if the original developer contacts them, ownership and administrative rights for the page will be transferred. This is how VIDRAFT reclaimed control of the
FINAL-Benchpage: they discovered the mirrored listing, emailed the contact address provided, and ModelScope transferred full management rights and completed an official verification process (indicated by a verified badge on the page).Ecosystem onboarding: After the handoff, VIDRAFT reported receiving a separate outreach from Alibaba Cloud proposing a collaboration — turning what started as passive mirroring into an active partnership conversation. ModelScope's interpreted goal: use mirrored content as a hook to recruit international developers into its ecosystem.
License compliance matters here. Not every model on Hugging Face can be freely mirrored. ModelScope's approach depends on the specific redistribution terms of each model's license. All Korean companies identified in this report indicated they do not consider the mirroring itself a license violation given their models' open-source terms.
Benchmarks & results
No specific quantitative benchmark numbers for FINAL-Bench itself are reported in this source article. The article focuses on the distribution and platform mechanics rather than evaluation results.
What is qualitatively described: FINAL-Bench is positioned as an AI evaluation benchmark developed and publicly released by VIDRAFT, significant enough to be indexed and mirrored by ModelScope as part of its outreach to non-Chinese developers.
How to try it
VIDRAFT's FINAL-Bench is publicly available on Hugging Face as the primary official channel. VIDRAFT confirmed Hugging Face remains their sole official public release platform.
The FINAL-Bench page also now exists on ModelScope with VIDRAFT holding verified ownership after the handoff process described above.
No OpenAI-compatible API endpoint, GitHub repository link, or pip install package for FINAL-Bench is mentioned in the source article. To access it:
- Search for
FINAL-Benchby VIDRAFT on huggingface.co - Search for the same on modelscope.cn where VIDRAFT now holds the verified page
FAQ
Q: If my model is on Hugging Face under an Apache 2.0 or MIT license, will ModelScope mirror it automatically?
A: Based on this reporting, it's plausible — ModelScope appears to continuously monitor Hugging Face and mirror models whose licenses permit redistribution with near-zero lag. Whether any specific model is targeted depends on factors not fully disclosed, but the pattern is clearly systematic rather than ad hoc.
Q: If I find my model on ModelScope without having put it there, what can I do?
A: According to VIDRAFT's experience, you can contact ModelScope via the email address listed on the Community Mirror page. ModelScope will then transfer page ownership and management rights to you and run a verification process. You'll receive an official verified badge once complete — and possibly a partnership inquiry from Alibaba Cloud afterward.
Q: Does claiming the ModelScope page mean I'm agreeing to a partnership with Alibaba Cloud?
A: Not automatically. VIDRAFT's account indicates that a partnership proposal came as a separate, subsequent outreach after the technical handoff was complete. The ownership transfer itself appears to be a distinct process from any commercial arrangement.
Originally reported by IT조선 (2026-10-05) — source article.
Top comments (0)