Decision AI, Explained: Why Hugging Face Built a Demo for Darwin-35B-ZTC-Vision
aimachinelearningcomputervisionopensource
TL;DR
Most AI models write. A decision AI chooses. Darwin-35B-ZTC-Vision takes up to four images and a question and returns a calibrated probability for every option in one forward pass. Hugging Face's open-source team liked it enough to build a free demo and offer it to us. VIDRAFT's ZTC models are #1 on the Hugging Face official Typed Decisions leaderboard (0.812) and on the S1MB decision-engine benchmark in English and Japanese.
What is a decision AI?
Many real tasks are choices from a fixed list: is there a defect, which photo is sharper, what quality level is this. A text model writes a long answer that someone has to parse. A decision model returns the probability of each option directly, so a system can act on it: approve above 0.9, send to a person between 0.3 and 0.7.
What is ZTC?
ZTC (Zero-Token Confidence) reads the model's internal state once, before any text is generated, and turns it into a probability per option. No extra tokens and no second model. That makes it fast and its numbers are calibrated, so they mean what they say.
| Item | Darwin-35B-ZTC-Vision |
|---|---|
| Size | 35B mixture of experts, about 3B active per token |
| Input | up to 4 images + question (yes/no, choice, or score) |
| Output | a calibrated probability per option |
| License | Apache 2.0 |
| Downloads | 8,940 (Hugging Face, 28 Oct 2026) |
Records
| Leaderboard | Result |
|---|---|
| Hugging Face official Typed Decisions | #1, Darwin-27B-ZTC-v5, 0.812 |
| S1MB decision engines, English | #1, Darwin-27B-ZTC-v9, Borda 90.38 of 197 models |
| S1MB, Japanese | #1, Darwin-27B-ZTC-v5, 84.57 (14 Oct 2026) |
| S1MB, vision | Darwin-35B-ZTC-Vision merged 19 Oct with all 20 required tasks; #1 on the published results at submission (Borda 91.00). The live board is authoritative. |
The Hugging Face demo
Apolinário of the Hugging Face open-source team built an interactive Gradio demo on free ZeroGPU hardware, including a custom low-RAM loader for the 69 GB checkpoint, and offered to transfer it to us. It is moving to the FINAL-Bench organization. This is open-source support, not a certification.
- Model: huggingface.co/FINAL-Bench/Darwin-35B-ZTC-Vision
- Demo: huggingface.co/spaces/hugging-apps/darwin-35b-ztc-vision
FAQ
Does it write captions? No. It only scores the options you give it.
Where is it useful? Defect inspection, quality grading, anomaly checks and any fixed-choice visual decision.
Is the demo free? Yes. It runs on Hugging Face ZeroGPU.
Are the leaderboard scores verified? They are the values on each public leaderboard; Hugging Face boards list publisher-reported scores.