VIDRAFT Claims Five Leaderboard Leads for Darwin-180B-RSI Reasoning Model
The AMW Read
Five reported leaderboard leads meaningfully raise VIDRAFT's profile among model developers, while downstream impact remains limited without independent workload and commercial evidence.
VIDRAFT Claims Five Leaderboard Leads for Darwin-180B-RSI Reasoning Model
Korean startup VIDRAFT says Darwin-180B-RSI ranks first on five Hugging Face leaderboards covering mathematical, scientific, general and multimodal reasoning. The company reports perfect scores on AIME 2026 and HMMT 2026, 94.44% on GPQA Diamond, 88.12% on MMLU-Pro and 79.48% on MMMU-Pro. According to AI Times, the open model builds on Qwen3.8-Flash-Next and uses a 180-billion-parameter mixture-of-experts architecture that activates 10 of 512 experts per request. VIDRAFT describes a roughly 260,000-token context window. These results and technical claims are reported by the company; the article does not provide an independent assessment of real-world performance.
The release puts VIDRAFT on the foundation-model player map as a model adapter aiming to compete on reasoning quality. Its stated method combines selective model merging with recursive self-improvement: the model generates solutions, filters for verified answers and retrains on those results. That extends the model adaptation and evaluation strategy highlighted in AI Market Watch’s September coverage of VIDRAFT’s Hugging Face feature. The reported leaderboard gains are notable for an open model built on an existing base, but benchmark rank alone does not establish a durable advantage in serving cost, reliability or enterprise deployment.
Builders evaluating Darwin-180B-RSI can test whether its reported reasoning gains hold on fresh, task-specific problems and measure latency and cost under their own workloads. Investors can look for repeatable improvements across model generations and evidence that VIDRAFT can turn its adaptation methods into sustained usage. Those checks matter more for commercial traction than five leaderboard positions on a single release.