Hugging Face Open-Source Team Donates a Demo for VIDRAFT's Darwin-35B-ZTC-Vision Decision AI — No GPU Required
TL;DR: VIDRAFT's
Darwin-35B-ZTC-Visionis a 35B-parameter Mixture-of-Experts vision model that outputs calibrated probabilities over predefined answer choices — not free-form text — using a single forward pass via its Zero-Token Confidence (ZTC) technique. The Hugging Face open-source team proactively built and gifted a ZeroGPU-powered interactive demo, meaning you can test it right now in a browser. Developers building classification, quality inspection, or anomaly-detection pipelines should pay attention.
What it is
Darwin-35B-ZTC-Vision is VIDRAFT's latest release in a line of Decision AI models — a model category designed specifically for structured judgment tasks rather than open-ended text generation.
Key characteristics:
- Architecture: 35 billion total parameters in a Mixture-of-Experts (MoE) structure, with approximately 3 billion parameters active per token during inference.
- Input: Up to four images plus a natural-language question per inference call.
- Output: A probability distribution over a fixed, user-defined answer set — yes/no, a single option from multiple choices, or a 1–5 grade scale — returned in a single computation step.
- License: Apache 2.0 (commercially usable, open weights).
- Model size on disk: Approximately 69 GB.
The model is published on Hugging Face under the FINAL-Bench organization and had logged 8,940 downloads as of October 28, 2026.
How it works
The defining technical idea is Zero-Token Confidence (ZTC). Instead of auto-regressively generating tokens to form an answer, ZTC reads the model's internal state before any output token is produced and derives a confidence score for each candidate answer in one pass.
Why this matters to engineers:
- No extra tokens generated — latency is bounded by a single forward pass, not by answer length.
- No external verifier model — the confidence signal comes from the base model itself.
- Structured output by design — probabilities sum to 1 across the predefined answer labels, making downstream threshold logic straightforward.
A practical workflow the source describes: if the returned probability for the positive class falls in an ambiguous range (e.g., 0.3–0.7), route the sample to a human reviewer; otherwise accept the automated decision. This makes it easy to build human-in-the-loop pipelines with a single configurable threshold.
Intended domains include manufacturing defect inspection, video/image quality assessment, and anomaly detection — anywhere fast, consistent judgment from a closed label set is required.
Benchmarks & results
All figures below are sourced directly from the press article. Rankings are point-in-time and may shift as new models are submitted.
| Benchmark | Model | Score | Rank (at time of report) |
|---|---|---|---|
| Typed Decisions (HF-certified) | Darwin-27B-ZTC-v5 | 0.812 | 1st |
| Decision Index (HF-certified) | Darwin-27B-ZTC | 63.03 | 1st |
| Decision Index Vision (HF-certified) | Darwin-35B-ZTC-Vision | 66.02 | 1st |
| S1MB English (197 models) | Darwin-27B-ZTC-v9 | Borda 90.38 | 1st (as of article date) |
| S1MB Japanese | Darwin-27B-ZTC-v5 | Borda 84.57 | 1st (as of Oct 14) |
| S1MB Vision | Darwin-35B-ZTC-Vision | Borda 91.00 | 1st at submission (Oct 19) |
VIDRAFT reports holding the most Hugging Face–certified benchmark #1 rankings of any single organization by their own count — 15 out of 53 certified benchmarks at the time of publication.
⚠️ Live leaderboard positions change as new models are submitted. Verify current rankings directly on Hugging Face.
How to try it
The Hugging Face open-source team member Apolinário built an interactive Gradio demo running on ZeroGPU (Hugging Face's free shared-GPU service) and transferred it to the VIDRAFT/FINAL-Bench account. The transfer was in progress at the time of the article.
Browser demo: Visit the FINAL-Bench organization on Hugging Face to find the Darwin-35B-ZTC-Vision Space. No GPU or API key needed — ZeroGPU handles compute.
Model weights: The weights are publicly available on Hugging Face under Apache 2.0. You can download them with:
huggingface-cli download FINAL-Bench/Darwin-35B-ZTC-Vision
Note: The Hugging Face open-source team developed a custom low-memory loader specifically to fit the ~69 GB model into the demo server's memory. If you self-host, factor in MoE memory management accordingly.
No GitHub repository URL or OpenAI-compatible API endpoint was mentioned in the source article at this time.
FAQ
Q: Why output probabilities instead of text — what's the engineering benefit?
A: Probability outputs over a closed label set are deterministic in structure, trivially parseable, and directly usable in threshold-based routing logic without any post-processing NLP. For production pipelines this eliminates an entire parsing layer compared to extracting a decision from free-form LLM output.
Q: Does the ~3B active parameter count mean I can run this on a single consumer GPU?
A: MoE models load all expert weights into memory even when only a subset activates per token, so the full ~69 GB of weights must fit in VRAM or be offloaded. The active-parameter count reflects compute cost per token, not memory footprint. Plan for multi-GPU or high-VRAM setups for local inference; the browser demo via ZeroGPU is the easiest zero-hardware path.
Q: Is commercial use permitted?
A: Yes. The model is released under the Apache 2.0 license, which permits commercial use, modification, and redistribution with attribution.
Originally reported by AI타임스 (2026-10-30) — source article.
Top comments (0)