A score of 0.81 puts the Korean AI startup back at the top of one of the most rigorous structured-decision leaderboards on Hugging Face.
TL;DR: VIDRAFT's Darwin-27B-ZTC-v5 has reclaimed the top position on the Hugging Face-certified Typed Decisions benchmark, scoring 0.81 on the accuracy metric as of October 16, 2026. The result places Darwin-27B-ZTC-v5 ahead of its predecessor, Darwin-27B-ZTC, which scored 0.74, and all other competing models on the public leaderboard. The benchmark evaluates structured probabilistic decision-making across real-world workflow scenarios including customer service, invoice processing, and security incident management.
VIDRAFT, the Korean Pre-AGI AI startup, has reclaimed the number-one spot on the Hugging Face Typed Decisions benchmark leaderboard as of October 16, 2026, with its latest model, Darwin-27B-ZTC-v5, achieving an accuracy score of 0.81. The result marks a significant milestone for the company, demonstrating measurable progress over its own prior submission and outpacing every other model currently listed on one of Hugging Face's officially recognized evaluation datasets.
VIDRAFT's Darwin-27B-ZTC-v5 topped the public leaderboard of the LocalLLaMA/typed-decisions dataset on Hugging Face, an officially designated benchmark focused on structured decision classification tasks. The model registered an accuracy score of 0.81, surpassing the previous Darwin-27B-ZTC, which had held a score of 0.74 on the same leaderboard.
The Typed Decisions benchmark is designed to evaluate how well a model handles probabilistic, multi-label classification across complex, realistic business workflows. The dataset spans five subsets: agent trace observability, customer service, invoice processing, and security incident management, with a total size in the 1K–10K range. Each row in the dataset includes structured fields for outcome, action, review necessity, risk, and urgency — each labeled with a confidence score and a set of class probabilities — making it a demanding test of both classification accuracy and calibration.
The leaderboard, which currently filters results by accuracy under the "Base only" setting, shows Darwin-27B-ZTC-v5 at the top with a score of 0.81. In second and third place respectively are Darwin-27B-ZTC (0.74, also from VIDRAFT) and empiriolabsai/aplomb-1 (0.74), followed by two models from ProKope-AI at 0.72 and 0.71. The previous leading model from the adaptive-classifier team scored 0.61, indicating a substantial performance gap between the top tier and the rest of the field.
Benchmark performance on Hugging Face's officially certified leaderboards carries weight in the AI research community because these evaluations are standardized, publicly reproducible, and not self-reported. Typed Decisions in particular is notable for testing something that many general-purpose language model benchmarks do not emphasize: structured, calibrated reasoning under realistic operational conditions.
Rather than measuring open-ended generation quality or raw knowledge recall, the Typed Decisions benchmark tests whether a model can correctly classify complex business states — such as whether a customer service interaction requires immediate human escalation, or whether a detected security event represents a critical risk — and do so with well-calibrated confidence estimates. This maps directly onto the kinds of decisions that enterprise AI systems are increasingly expected to make autonomously or semi-autonomously in production environments.
For VIDRAFT, reclaiming the top position with a meaningful score improvement over its own prior model signals that the company's iterative development process is producing real gains on these structured-decision tasks. The fact that VIDRAFT now occupies both first and second place on the leaderboard underscores the distance between its current capabilities and those of competitors in this specific evaluation domain.
The benchmark's Apache 2.0 license and public availability on Hugging Face also mean that these results are fully auditable by the broader research community — adding credibility to the ranking beyond any single organization's claims.
Q: What is the Hugging Face Typed Decisions benchmark?
A: Typed Decisions (LocalLLaMA/typed-decisions) is an officially recognized Hugging Face benchmark dataset that tests AI models on structured, multi-label probabilistic classification tasks drawn from real-world business workflows such as customer service, invoice processing, and security incident management. Models are evaluated on accuracy and confidence calibration.
Q: What score did VIDRAFT's Darwin-27B-ZTC-v5 achieve, and how does it compare to other models?
A: Darwin-27B-ZTC-v5 scored 0.81, placing it first on the leaderboard. The next closest models — Darwin-27B-ZTC and empiriolabsai/aplomb-1 — each scored 0.74, while other competitors ranged from 0.61 to 0.72.
Q: Why is reclaiming the top spot on this benchmark significant for VIDRAFT?
A: The Typed Decisions benchmark focuses on calibrated, structured decision-making under realistic conditions — a capability central to enterprise AI deployment. Occupying both first and second place on a publicly auditable, Hugging Face-certified leaderboard provides independently verifiable evidence of VIDRAFT's progress in this domain.
Source: FINAL-Bench (2026-10-16) — original article