VIDRAFT Releases Darwin-180B-RSI… Achieves Five First-Place Rankings Including Perfect Score on AIME2026

VIDRAFT has released Darwin-180B-RSI, which achieved first place in five categories on the official Hugging Face leaderboard by modifying approximately 0.02% of the parameters of the open model Qwen3.8-Flash-Next (180B MoE). The core technique is Recursive Self-Improvement (RSI), where the model solves problems and re-trains itself by grading against verifiable correct answers without human labels.

Benchmark Performance

The model achieved the following records on the official Hugging Face leaderboard:

• AIME 2026: 100.0 points (First perfect score in leaderboard history)
• HMMT Feb 2026: 100.0 points (First perfect score in leaderboard history)
• GPQA Diamond: 94.44 points (Approximately 30 percentage points higher than the PhD average)
• MMLU-Pro: 88.12 points
• MMMU-Pro vision: 79.48 points

The measurement conditions were applied identically to all benchmarks: thinking budget of 131,072 tokens, temperature 1.0, top_p 0.95, top_k 20, and precision bf16.

Technical Features and Efficiency

The model architecture modified only the attention and shared experts of the parent model, while keeping the 512 routing experts and the vision encoder unmodified. This maintained the same accuracy as the parent model while reducing inference length by approximately 11%, thereby lowering GPU inference costs. Additionally, the ZTC (Zero-Token Confidence) technique, which estimates the probability of a correct answer from internal states before generation, was applied to pre-filter potential errors in AI agents.

Release and Usage

The weights are released on Hugging Face under the Qwen Community License 1.0, allowing anyone to download and verify them. Serving uses vLLM and requires tensor parallelism of 8 and Expert Parallel. VIDRAFT stated that achieving world number one by evolving only parts of an open model, without training new ultra-large models from scratch, proves the technological competitiveness of Korean startups.