Not tested by us yet. SingVL_S1_4B is open weights only, and it needs a custom loading class and a GPU, so we have not run it. Everything below comes from the Hugging Face model card. We'll update this post when we have run it next to a chat model acting as a classifier.
SingVL_S1_4B is a decision model from inclusionAI: instead of writing text, it reads a question plus your text, JSON or up to four images and returns probabilities. It is built on Qwen3.5 and sits at about 5B parameters in BF16 Safetensors (the "4B" in the name is not the stored size). It is for teams that want a classifier or grader they can host themselves, including on images, and not a general chat model.
| SingVL_S1_4B | |
|---|---|
| Maker | inclusionAI |
| Base model | Qwen3.5 |
| Size | About 5B, BF16 Safetensors |
| Inputs | Text, JSON, up to 4 images, plus a question |
| Outputs | Probabilities, one forward pass per question |
| Access | Open weights on Hugging Face |
| License | Not stated on the model card |
| Price | None; you host it |
What it does
The model card describes three question types, each answered in a single Transformers forward pass.
- Choice: you supply candidate options and get a probability for each one.
- Noul: you supply a proposition and get the probability that it is true.
- Score: you pick 2 to 10 ordered levels and get a probability for each, plus an expected level.
That covers most of what people use an LLM call for inside an agent: routing a request, checking a yes/no condition, grading an answer on a scale. The difference is that you read a probability, so you can set your own threshold instead of parsing a sentence. Input limits are 65,536 tokens for text, or 10,000 tokens when images are included.
You can load it with Transformers through AutoModelForMultimodalLM, use a SingVLS1 class, or run the bundled FastAPI server at /v1/systemone.
The numbers, and who reported them
All scores below are from the model card, not from us. They use a benchmark called Public Decision Index 0.3, and we have not verified the benchmark itself or reproduced a single score.
| Panel | SingVL_S1_4B | Comparison models on the card |
|---|---|---|
| Text | 69.76 | 60.70, 67.8, 60.1 |
| Vision | 73.94 | 67.90, 72.5 |
The card does not name the comparison models in the part we could read, and the vision gap over the closest listed model is about 1.4 points. The same card lists weak results on some tasks: 12.77 accuracy on HLE and 8.33 macro-F1 on SGD/SGD-X. Read the table as "competitive on its own index", not as proof that it beats a chat model on your data.
Coverage of the release is thin. The only other write-ups we found come from a single router blog, and they contradict each other on the size, so we used the model card for every figure here.
How it compares on paper
Dedicated decision models are a new category, and most of them are hard to get. Kakao's Kanana-2-v-decide is the closest relative: an open decision model that also takes images. Ingenix's Thalamus Decision is open but aimed at biology. boost.ai's Presis and Salesforce's Koa Decisions are private components you cannot download.
What stands out about SingVL_S1_4B is the combination: open weights, three question types including graded scores, and image input in the same model. Its limit is the same as any weights-only release. You carry the serving, the GPU cost and the licence question, since the card states no licence.
Who should care
- Teams that already self-host and classify images or documents. A single forward pass that returns calibrated-looking probabilities is easier to threshold and cheaper to run than a generative call.
- Anyone building graders or routers. The score type gives you an expected level without asking a chat model to emit a number.
- Skip it if you need a hosted API, a clear commercial licence or a vendor SLA. None is available today.
What we will test
When we run it, we would put it next to a cheap and a mid-priced chat model acting as classifiers, on one shared set of 30 to 60 items with control items that have an obvious answer. We would ask yes/no questions in both directions to see whether the probability flips when the wording is mirrored, then report accuracy, median latency and cost per 1,000 decisions. Until then, treat every score above as the maker's own.