When the platform comes to you first, it signals something bigger than a benchmark win.
TL;DR: VIDRAFT's vision-based Decision AI model, Darwin-35B-ZTC-Vision, received an unsolicited interactive demo built and donated by the Hugging Face open-source team in late October 2026. The demo runs on Hugging Face's ZeroGPU service, requiring no local GPU from users. VIDRAFT reports holding more top-1 rankings across Hugging Face-certified benchmarks than any other organization by its own count.
Korean AI startup VIDRAFT announced on October 30, 2026 that the open-source team at Hugging Face headquarters had independently created and donated a hands-on demo for its publicly released vision Decision AI model, Darwin-35B-ZTC-Vision — without any request from VIDRAFT to do so.
A member of Hugging Face's open-source team, Apolinário, posted on the model's public page on October 27, 2026, congratulating VIDRAFT on the release and expressing appreciation for the work. He described the model as impressive and announced that he had built an interactive demo app running on ZeroGPU — Hugging Face's free GPU hosting service — and offered to transfer it to VIDRAFT's account. VIDRAFT accepted, and the demo is now being migrated to the FINAL-Bench organization on Hugging Face.
What makes this notable is the direction of initiative. VIDRAFT did not pitch the idea or ask for help. The Hugging Face team discovered the model independently, evaluated it, and decided to build the demo on their own. To handle the model's roughly 69-billion-byte footprint on demo server memory, the Hugging Face team also developed a custom low-memory loader — enabling any user to test the model directly in a web browser, no GPU hardware required.
Darwin-35B-ZTC-Vision is built around what VIDRAFT calls a "Decision AI" paradigm. Rather than generating free-form text responses, the model takes up to four images and a question as input, then outputs a probability distribution across a fixed set of answer options — yes/no, a choice from multiple options, or a 1-to-5 rating — in a single inference pass. For instance, feeding the model an image of a manufactured part and asking whether a defect is present might return a result like "Yes: 0.82, No: 0.18." The model does not narrate its reasoning; it delivers structured probability scores directly.
The core technology enabling this is ZTC, or Zero-Token Confidence, which reads the model's internal state once before any output tokens are generated, deriving a confidence score for the judgment without producing additional tokens or invoking a separate verification model. This structure makes it straightforward to embed threshold logic into enterprise workflows — for example, routing any prediction between 0.3 and 0.7 probability to a human reviewer automatically.
The model uses a Mixture of Experts architecture with 35 billion total parameters, activating approximately 3 billion parameters per token during inference. It is released under the Apache 2.0 license, allowing commercial use. As of October 28, 2026, it had been downloaded 8,940 times on Hugging Face.
The unsolicited demo contribution from Hugging Face is an unusual form of external validation. Platform teams at that scale typically support models reactively, if at all. Choosing to build tooling around Darwin-35B-ZTC-Vision — and developing a specialized memory loader to make it accessible for free — suggests the Hugging Face team sees the Decision AI category as technically meaningful and worth surfacing to a broader audience.
VIDRAFT's ZTC series has also accumulated a string of first-place finishes on Hugging Face-certified benchmarks. On the Typed Decisions benchmark, Darwin-27B-ZTC-v5 ranks first with a score of 0.812. On Decision Index, Darwin-27B-ZTC holds first place at 63.03 points. Darwin-35B-ZTC-Vision leads Decision Index Vision with 66.02 points. On the System One Mosaic Benchmark (S1MB), Darwin-27B-ZTC-v9 ranks first among 197 models in the English edition with a Borda score of 90.38, while Darwin-27B-ZTC-v5 leads the Japanese edition with 84.57 points as of October 14. In the S1MB Vision edition, Darwin-35B-ZTC-Vision completed all 20 required tasks without failure and was officially listed on October 19, posting a Borda score of 91.00 at the time of submission. Rankings may shift as additional models are added to the leaderboard. VIDRAFT states that across 53 Hugging Face-certified benchmarks, it holds first place in 15 — more than any other organization by its own count.
The practical applications targeted include manufacturing defect inspection, video quality assessment, and anomaly detection — settings where fast, consistent, structured judgment is more useful than verbose natural language output.
Q: What is Darwin-35B-ZTC-Vision and what makes it different from a standard AI model?
A: Darwin-35B-ZTC-Vision is a Decision AI model that takes images and a question as input and returns probability scores across fixed answer choices, rather than generating free-form text. Its ZTC technology derives confidence scores from the model's internal state in a single inference pass, with no extra token generation.
Q: Who built the Hugging Face demo for VIDRAFT, and why is that significant?
A: A member of Hugging Face's own open-source team, Apolinário, built the demo independently and offered to donate it to VIDRAFT — an unsolicited gesture that signals platform-level recognition of the Decision AI category.
Q: Do users need a GPU to try the Darwin-35B-ZTC-Vision demo?
A: No. The Hugging Face team developed a custom low-memory loader so the model runs on Hugging Face's free ZeroGPU service, making it accessible through any web browser without local GPU hardware.
Source: AI타임스 (2026-10-30) — original article