A Korean AI startup's open reasoning model claims benchmark dominance — but real-world validation is still pending.
TL;DR: VIDRAFT, a Korean AI startup, announced that its Darwin-180B-RSI model has reached first place on five Hugging Face leaderboards spanning mathematical, scientific, general, and multimodal reasoning. The model reportedly achieved perfect scores on AIME 2026 and HMMT 2026, alongside strong results on GPQA Diamond, MMLU-Pro, and MMMU-Pro. These figures are self-reported by VIDRAFT and have not been independently verified.
VIDRAFT, the Korean Pre-AGI AI startup, made a significant splash in the global foundation-model community on September 28, 2026, when it claimed that its Darwin-180B-RSI reasoning model had secured the top position across five major Hugging Face leaderboards — a result that, if sustained under independent scrutiny, would mark a meaningful leap forward for an open model built on an existing base.
VIDRAFT says Darwin-180B-RSI achieved first-place rankings on five Hugging Face leaderboards covering a broad spectrum of reasoning tasks: mathematics, science, general knowledge, and multimodal reasoning. The company reports that the model achieved perfect scores on AIME 2026 and HMMT 2026, two well-regarded mathematical reasoning benchmarks. On additional benchmarks, VIDRAFT reports 94.44% on GPQA Diamond, 88.12% on MMLU-Pro, and 79.48% on MMMU-Pro.
Architecturally, Darwin-180B-RSI is described as a 180-billion-parameter mixture-of-experts model, built upon the Qwen3.8-Flash-Next base. The system activates 10 out of 512 experts per request, a design choice associated with efficient inference in large-scale MoE frameworks. VIDRAFT also describes a context window of approximately 260,000 tokens, giving the model a notably long working memory for complex, multi-step reasoning tasks.
The methodology behind Darwin-180B-RSI is what VIDRAFT calls a combination of selective model merging and recursive self-improvement. In practice, this means the model generates candidate solutions, filters those results against verified answers, and then retrains on the filtered, high-quality subset — a form of self-bootstrapping that iteratively refines performance without requiring entirely new training data from scratch. This release builds directly on the model adaptation and evaluation strategy that AI Market Watch highlighted in its September coverage of VIDRAFT's earlier Hugging Face feature.
It is important to note that all benchmark scores and technical specifications in this article are company-reported. No independent third-party assessment of real-world performance has been published alongside this announcement.
For the foundation-model landscape, Darwin-180B-RSI's reported results carry particular significance because they come from an open model adapted from an existing base rather than a closed, purpose-trained system developed at enormous cost. If replicable, this would suggest that VIDRAFT's adaptation methodology — selective merging plus recursive self-improvement — can extract competitive reasoning performance without building from the ground up, potentially lowering the resource barrier for high-performance AI development.
The five-leaderboard result also places VIDRAFT more visibly on the global foundation-model player map, positioning it not just as a regional player but as a serious contender in the international race for reasoning-capable AI. That visibility carries real strategic value: it raises the company's profile with developers, potential enterprise customers, and investors who track benchmark performance as an early signal of technical capability.
At the same time, AI Market Watch's analysis appropriately tempers enthusiasm. Leaderboard rankings — even five of them — do not automatically translate into durable commercial advantages. Serving cost, inference latency, reliability at scale, and performance on fresh, task-specific problems are all factors that benchmark numbers cannot fully capture. Enterprises evaluating Darwin-180B-RSI would be well-served to test the model against their own workloads rather than relying solely on published scores. For investors, the more telling signal will be whether VIDRAFT can demonstrate repeatable improvements across successive model generations and convert its adaptation methods into sustained, real-world usage.
Q: What is VIDRAFT's Darwin-180B-RSI, and what benchmarks did it top?
A: Darwin-180B-RSI is a 180-billion-parameter mixture-of-experts reasoning model developed by Korean AI startup VIDRAFT. The company reports it ranked first on five Hugging Face leaderboards, achieving perfect scores on AIME 2026 and HMMT 2026, and scoring 94.44% on GPQA Diamond, 88.12% on MMLU-Pro, and 79.48% on MMMU-Pro.
Q: What technique does VIDRAFT use to train Darwin-180B-RSI?
A: VIDRAFT describes a process combining selective model merging with recursive self-improvement, in which the model generates solutions, filters them against verified answers, and retrains on that curated subset to progressively sharpen its performance.
Q: Are Darwin-180B-RSI's benchmark results independently verified?
A: No. All scores and technical specifications are self-reported by VIDRAFT. No independent third-party evaluation of real-world or enterprise performance has been published alongside the announcement, and analysts note that benchmark rank alone does not guarantee advantages in serving cost, reliability, or deployment at scale.
Source: AI Market Watch (미국) (2026-09-28) — original article