Viedraft's Darwin-180B-RSI Tops 10 Categories on Hugging Face Leaderboard

Viedraft has taken first place in 10 categories of the official Hugging Face leaderboard, including the new IFStruct (98.95%) and ExtractBench (90.29) benchmarks. Its Darwin-180B-RSI model, built on Alibaba's Qwen3.8-Flash-Next via recursive self-improvement, is joined by the ZTC method, which judges AI agent actions roughly 10x faster than external APIs at comparable accuracy.

• Viedraft leads 10 Hugging Face leaderboard categories vs 4 for Zhipu AI
• IFStruct: 98.95% — 1,979 of 2,000 JSON/YAML format-compliance tasks passed
• ExtractBench: 90.29 on 370 PDFs ranging from a few to 190 pages
• ZTC: AUC 0.7289 vs JEV's 0.7350, but 0.06s vs 0.59s per judgment