A Korean AI startup's 180B model outscores trillion-parameter rivals on two major benchmarks
TL;DR: VIDRAFT's Darwin-180B-RSI has been listed as the number-one model on both the MMLU-Pro and GPQA leaderboards maintained by Russian AI benchmark aggregator AIhoster, scoring 88.1% and 94.4% respectively. The rankings, updated on October 19, 2026, place Darwin-180B-RSI ahead of models from DeepSeek, Qwen, MiniMax, and Moonshot AI — several of which carry parameter counts many times larger. AIhoster pulls scores directly from public model cards without recalculating them.
VIDRAFT's Darwin-180B-RSI claimed the top position on two independent benchmark leaderboards published by AIhoster, a Russian open-model evaluation aggregator, as of its October 19, 2026 update. The listings place the Korean AI startup's 180-billion-parameter model above a competitive field that includes models with parameter counts reaching into the trillions.
AIhoster maintains separate leaderboards for distinct benchmark tasks, explicitly choosing not to merge scores into a single composite number — a methodology the site notes reflects the fact that each benchmark uses different datasets, evaluation conditions, and measurement approaches. Scores are sourced directly from public leaderboard tables and matched by exact repository ID.
MMLU-Pro — Rank #1 (88.1%)
On the MMLU-Pro leaderboard, which tests challenging multiple-choice questions across a broad range of knowledge domains, Darwin-180B-RSI recorded an accuracy of 88.1% across all 12,032 test items. The evaluation used a single-sample configuration. The result placed Darwin-180B-RSI ahead of MiniMaxAI's MiniMax-M2.1 (229B parameters, 88.0%), InternLM's Intern-S2-Preview (36.1B, 88.0%), Qwen3.5-397B-A17B (403B, 87.8%), and DeepSeek-V4-Pro (1.6 trillion parameters, 87.5%), among others.
GPQA — Rank #1 (94.4%)
On the GPQA Diamond leaderboard — which targets expert-level questions in physics, chemistry, and biology — Darwin-180B-RSI scored 94.4% across all 198 items, evaluated using majority voting over up to 16 samples with a generous thinking budget. This result edged out Moonshot AI's Kimi-K3 (2.8 trillion parameters, 93.5%), as well as VIDRAFT's own Darwin-397B-ZTC (404B parameters, 93.4%) in third place, and Ornith-1.5-397B (403B, 92.8%) in fourth. Notably, VIDRAFT's Darwin-398B-JGOS also appeared in the GPQA top ten at ninth place with 90.9%, giving the Korean startup multiple entries in the upper tier of the chart.
The benchmark results are significant for several reasons. First, Darwin-180B-RSI achieves these scores at 180 billion parameters — a figure that, while not small, is dwarfed by several of the models it outranks. DeepSeek-V4-Pro, for example, carries 1.6 trillion parameters, roughly nine times the size of Darwin-180B-RSI, yet scores 0.6 percentage points lower on MMLU-Pro. On GPQA, Kimi-K3 has 2.8 trillion parameters — more than fifteen times larger — and still trails Darwin-180B-RSI by nearly a full percentage point.
This parameter-efficiency narrative is increasingly central to how frontier AI labs differentiate themselves. A smaller model that outperforms much larger ones carries practical advantages: lower inference cost, reduced hardware requirements, and easier deployment at scale.
Second, having a third-party aggregator — one that explicitly avoids recalculating or normalizing scores and instead links directly to original model cards — list Darwin-180B-RSI at the top of both tables lends a degree of independent credibility to the results. AIhoster's methodology is conservative by design, which makes a number-one placement under those conditions meaningful.
Third, VIDRAFT's presence across multiple GPQA slots (ranks 1, 3, and 9 with three distinct Darwin variants) suggests a broad and active model development pipeline rather than a single one-off result.
Q: What scores did VIDRAFT's Darwin-180B-RSI achieve on MMLU-Pro and GPQA?
A: Darwin-180B-RSI scored 88.1% on MMLU-Pro across all 12,032 test items and 94.4% on GPQA Diamond across all 198 items, earning first place on both leaderboards as listed by AIhoster.
Q: How does Darwin-180B-RSI compare to larger models like DeepSeek-V4-Pro or Kimi-K3?
A: Darwin-180B-RSI outperformed both on their respective leaderboards — despite DeepSeek-V4-Pro being roughly nine times larger (1.6 trillion parameters) and Kimi-K3 being more than fifteen times larger (2.8 trillion parameters).
Q: Is AIhoster's ranking methodology independent?
A: Yes. AIhoster explicitly states it does not recalculate scores; it aggregates results directly from public benchmark tables and matches models by exact repository ID, without creating an artificial composite score.
Source: AIhoster (러시아) (2026-10-18) — original article