Darwin-35B-ZTC-Vision

Darwin-35B-ZTC-Vision is a vision decision model. It does not generate text. Given one or more images, a state and a typed question, it returns a calibrated probability for every option in one forward pass: yes or no (noul), one of several options (choice), or a level on an ordered scale (score). It handles up to four images per request, for example "which of the two photos has the higher quality" or "is there a defect on this part".

It is a 35B mixture-of-experts model (about 3B active parameters per token) built on Darwin-35B-A3B-Opus with the Darwin ZTC decision readout. It is the vision member of the ZTC series; the text member is Darwin-27B-ZTC-v9.

Results

S1MB (System One Mosaic Benchmark), vision-v1

All 20 required vision benchmarks complete (the optional Box tasks are not supported), zero failed decisions. Merged by the maintainer on 2026-10-19 (PR #76). At submission it ranked first (Borda 91.00) on the published vision results of 2026-10-17 (16 complete models) with this model added, recomputed with the leaderboard's own Borda method; that table is below. More models have entered since: on the published results of 2026-10-19 (32 complete models) the same recomputation places it fifth with 81.80. The live vision leaderboard is authoritative.

Rank (2026-10-17 data)	Model	Borda score	Mean baseline-adjusted score
1	Darwin-35B-ZTC-Vision	91.00	73.98
2	Jev 1.14 (Preview)	90.00	74.64
3	Perplexity Decider Flash 1.2	77.50	67.35
4	Imajev 4B	67.83	69.29
5	Clef-Flash	63.50	63.70

Selected benchmarks (baseline-adjusted score)

Benchmark	Darwin-35B-ZTC-Vision	Jev 1.14 (Preview)
tallyqa-score (counting)	0.720	0.546
visa-noul (anomaly detection)	0.603	0.467
koniq-score (image quality)	0.196	0.143
cvbench-relation-choice	0.965	0.944
pope-noul (object hallucination)	0.793	0.767
nights-choice (perceptual similarity, 3 images)	0.886	0.866
vsr-noul (spatial relations)	0.775	0.768
generalization-vision-easy-choice	0.976	0.967

Training and data

Full-weight fine-tuning with a decision readout on public train splits and our own synthetic data:

• COCO train2017 images (licensed subset) for verification, counting, relation and multi-hop spatial questions computed from the annotations, plus synthetic quality degradations rated on a five-level scale.
• VisA normal images from the official train split (CC BY 4.0) with synthetic surface defects, plus procedural textures.
• Public train splits of text and chart decision datasets, and the text decision data of the Darwin ZTC series.

No S1MB question text was used for training.

Usage

Python 3.12, a CUDA GPU with about 75 GB of free memory (BF16 weights), and the code shipped in this repository (autojev/).

import sys
from pathlib import Path
from huggingface_hub import snapshot_download

path = Path(snapshot_download("FINAL-Bench/Darwin-35B-ZTC-Vision"))
sys.path.insert(0, str(path))
from autojev.model import DecisionModel, answer

model = DecisionModel(checkpoint=path).eval()
question = {
    "type": "noul",
    "instructions": "Determine whether image_1 shows a defect such as a scratch, crack, hole or stain.",
    "criteria": {"true": "A defect is present", "false": "The surface is normal"},
}
row = {"state": {"images": ["image_1"]}, "images": ["part.jpg"], "question": question}
print(answer(question, model.predict([row])[0]))

decide_server.py serves POST /v1/systemone (images as data URLs, up to four per request):

ZTC_MODEL=/path/to/snapshot PORT=8410 python decide_server.py

s1mb_adapter/ztc_http.py is the S1MB evaluator adapter we used to measure the results above against this server.

License

Apache-2.0 for the weights and code in this repository. Training images keep their own licenses (COCO: per-image Flickr licenses and CC BY 4.0 annotations; VisA: CC BY 4.0).