Darwin-180B-RSI: VIDRAFT's Recursive Self-Improvement MoE Model Claims Five Leaderboard Tops
Darwin-180B-RSI: VIDRAFT's Recursive Self-Improvement MoE Model Claims Five Leaderboard Tops TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has released Darwin-180B-RSI — a 180-billion-parameter mixture-of-experts reasoning model built on Qwen3.8-Flash-Next that combines selective model merging with recursive self-improvement. The model reportedly tops five Hugging Face leaderboards across math,…
VIDRAFT, a Korean Pre-AGI AI startup, has released Darwin-180B-RSI, a 180-billion-parameter mixture-of-experts reasoning model. This model combines selective model merging with recursive self-improvement, claiming five leaderboards tops on Hugging Face. Darwin-180B-RSI's architecture is built on Qwen3.8-Flash-Next, featuring 180 billion total parameters and 512 experts, with 10 experts activated per inference request.
The model uses a sparse expert activation design, which keeps the effective compute per forward pass lower than a dense 180B model. VIDRAFT describes the model as a high-performance model adaptation rather than a ground-up pretrain. Darwin-180B-RSI's self-improvement works by generating candidate solutions, evaluating them against ground-truth answers, and retaining only the correct solutions for retraining in subsequent iterations.
The model claims perfect scores on five Hugging Face leaderboards, including AIME 2026, HMMT 2026, GPQA Diamond, MMLU-Pro, and MMMU-Pro. However, these scores are self-reported and have not been independently validated.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
