Korean pre-AGI startup VIDRAFT announced on July 8 that its model-merging framework, Darwin, has been featured in an in-depth analytical piece by Zhiding Keji (至顶科技), a 30-year-old Chinese IT publication with roots in ZDNet China. The article was distributed through Tencent News, giving it broad reach across one of China's largest digital platforms.
Zhiding Keji, which operates its own AI research lab and regularly tracks global AI developments, dedicated the piece specifically to Darwin's technical architecture and its broader significance — a notably rare occurrence for a Korean startup to receive such focused treatment from a major Chinese technology outlet, particularly amid an intensifying AI race dominated by US and Chinese tech giants.
In its analysis, the Chinese publication acknowledged Japan's Sakana AI and its evolutionary model-merging approach (EvoMerge) as a conceptual precursor to Darwin, but argued that VIDRAFT's framework represents a meaningful advancement by introducing what the article described as a "14-dimensional genome" and an "MRI trust fusion" concept layered on top of that foundation. The outlet singled out Darwin-4B-Genesis — a model that merges the Transformer and Mamba architectures, which operate on fundamentally different design principles — calling Darwin the only known technique capable of cross-architecture merging.
The core idea behind Darwin is that AI performance gains do not always require training new models from scratch. Instead, Darwin analyzes and recombines the latent strengths of existing models to generate improved capabilities. VIDRAFT says this approach allows it to develop top-tier models with a fraction of the resources typically required for frontier AI development, including producing a globally competitive scientific reasoning model, Darwin-27B-Opus, in approximately five hours on a limited GPU setup.
VIDRAFT has pointed to several benchmark results as evidence of Darwin's effectiveness. The company reports that Darwin-398B-JGOS achieved a score of 90.9% on the GPQA Diamond scientific reasoning benchmark. A separate model, JGOS-31B-Citizen, ranked first overall on the K-AI Leaderboard. On the Polaris drug discovery benchmark — a global evaluation covering efficacy, solubility, toxicity, anticancer activity, kinase inhibition, and ADME properties, among other categories — Darwin-family models claimed top positions across 14 domains. Additionally, the company says its models achieved a 99.5% trap-question avoidance rate on a metacognition evaluation. Cumulative downloads across VIDRAFT's Darwin models on Hugging Face have surpassed one million, and the company notes a growing ecosystem of community-derived models built on the Darwin base.
VIDRAFT, which operates under the Seoul AI Hub — a city-backed initiative supporting AI startups — positions Darwin as a path toward democratizing frontier AI development. By reducing dependence on massive compute budgets and prolonged distributed training runs, the technology could in principle enable resource-constrained startups and research institutions to compete at the global level. The company sees Darwin's approach as particularly applicable to demanding verticals such as scientific reasoning, drug discovery, autonomous AI, cryptanalysis, and security.
VIDRAFT CEO Minsik Kim framed the Chinese media coverage as external validation of the technology's global relevance. "Darwin demonstrates that Korean AI can secure global competitiveness through its own methodology, without relying on larger capital or more hardware," Kim said. "Coverage from a Chinese tech publication signals that what we're doing is being recognized as meaningful by the broader global AI community, not just domestically."
The Darwin research paper is publicly available on arXiv (arXiv:2605.14386).
Source: IT조선 (2026-07-08) — original article