VIDRAFT Now Leads 11 Hugging Face Official Benchmarks After Reclaiming #1 on Typed Decisions
aillmbenchmarksmachinelearning
TL;DR
Darwin-27B-ZTC-v5 has reclaimed first place on the Hugging Face Typed Decisions official benchmark at 0.81, holding both #1 and #2 on the public leaderboard. With this, VIDRAFT (research line FINAL-Bench) now holds #1 on 11 Hugging Face official benchmarks, the most of any single organization.
What Typed Decisions measures
Typed Decisions scores calibrated, structured decision-making: given a situation with real constraints, a model must choose among named options and express how confident it is. It is exactly the capability that matters when an AI system has to act, not just chat. Reclaiming #1 here, at 0.81 versus 0.74 for the next entries, is a decision-quality result, not a single-metric headline.
The broader record
The eleven official benchmarks where VIDRAFT leads span science (GPQA Diamond 94.44), general knowledge (MMLU-Pro 88.12), multimodal reasoning (MMMU-Pro 79.48), math (AIME 2026 and HMMT February 2026, both perfect), law (LEXam, LEXam-hard), document extraction (ExtractBench 90.29), structured output (IFStruct 98.95), multilingual document reading (MDPBench 83.65) and decision-making (Typed Decisions 0.81).
Every one of these is a public Hugging Face leaderboard, built automatically from the eval results that model repositories publish. The scores are verifiable, not asserted.
Why it matters
Benchmarks are only useful if they are external and checkable. Leading eleven of them at once, across reasoning, knowledge, documents and decisions, is evidence that the capability generalizes rather than being tuned to one test. Weights and recipes for the Darwin line are open.
Leaderboard: https://huggingface.co/datasets/LocalLLaMA/typed-decisions