SingVL_S1_4B: The 4.54B Decision Model inclusionAI Shipped Without a Word

At 03:58 UTC on November 11, 2026, a repository called SingVL_S1_4B appeared under the inclusionAI organisation on Hugging Face. At 05:38 UTC, someone edited its README. By the time this was written there was no blog post from the lab, no arXiv paper, no GitHub release, no entry in any inference provider's catalogue, and the download counter sat at zero. The model card spells the name with spaces — Sing VL S1 — and describes it as a multimodal decision model built on Q​wen3.5. That is the entire public record. Everything below is read off the files in that repository, and everything that reads like a claim is labelled as one.

What "decision model" means here, precisely

This is not a chat model with vision bolted on. You do not send it a prompt and read a paragraph back. You send it a state — text, JSON, or up to four images — plus a dictionary of named questions, and each question declares its type and its candidate answers. The model runs one forward pass and returns a probability for every candidate.

Three question types exist, and the names are the model's own:

• Choice — pick one of the supplied options; the response carries a probability per option.

• Noul — a yes/no proposition; the response is the probability that it is true. The spelling is odd and it is what the card uses.

• Score — rate the input on 2 to 10 ordered levels; the response carries each level's probability plus an expected level computed as the sum of zero-based index times probability.

So a single request can ask "which colour is named", "is the named colour red", and "how clearly does the text identify the colour as red", and get three answers with confidences from one pass. The card's own example does exactly that, then repeats it across two images to ask which one is red.

The architectural consequence is that the output vocabulary is not a vocabulary. There is no sampling, no temperature, no token stream to parse and no JSON mode to coax. If your pipeline currently ends in a regex over a model's prose to extract a label, this is the shape of model that removes the regex. If your pipeline needs the prose — a summary, a rationale, a tool call — this model has nothing to give you, because it never generates one.

The numbers, and whose numbers they are

The card publishes a table headed "Public Decision Index 0.3". Sing VL S1 scores 69.76 on the Text panel and 73.94 on the Vision panel. The comparison columns are TypeSafe Jev 1.14 at 60.70 on text with no vision score published, StartLux-Decision v1.1 4B at 67.8 text and 67.90 vision, and Perplexity Decider Flash v1.2 (4B) at 60.1 text and 72.5 vision.

Read that as a self-report. Every figure in it comes from the lab that trained the model, on a leaderboard whose scoring kit is public but whose private half is not, and no third party has re-run any of it. The Decision Index's own reproduction kit — the decision-index project that rebuilds the public suite and scores submitted runs — shows no SingVL submission anywhere in its issue or pull request history as of this morning. The card does include the submission instructions, which is a fair signal of intent and not evidence of a result.

The per-dataset breakdown is more useful than the headline, because it shows where the aggregate comes from. On the eleven vision datasets the strongest rows are InfographicVQA at 96.31, Winoground at 93.12 and the CORD-plus-FUNSD key-information set at 95.82. The weak rows are as informative: MMMU-Pro vision lands at 50.17 non-skill and 42.85 with the skill prompt, R-Bench-M at 49.77 and 39.72, and hateful-meme moderation at 71.64 falling to 43.28. That is a model that reads documents and layouts well and reasons about hard academic figures poorly — a profile consistent with a 4.5-billion-parameter backbone doing classification rather than deliberation.

On the forty-three text datasets, BFCL at 96.81 and CLINC150 at 97.03 are the standouts; HLE at 12.77, ChessBench at 26.88 and the SGD dialogue set at 8.33 are the floor. The card marks several rows with a dagger to say the index reports them separately rather than folding them in, which is the kind of disclosure that makes the rest of the table easier to trust.

What is actually in the repository

Two safetensors shards totalling about 9.1 GB, which at BF16 works out to roughly 4.54 billion parameters — the repository name says 4B, the Hugging Face sidebar rounds to 5B, and the tensor total is the number in between. A configuration declaring the Q​wen3.5 conditional-generation architecture with a hidden size of 2560. A 7.7 KB chat template that handles image and video tokens. And an inference/ directory containing the thing that makes any of this usable: a Python wrapper, a FastAPI server, a benchmark script and two example images, one blue and one red.

The setup is specified down to the patch version: Python 3.12, hardware-matched PyTorch and torchvision, requirements installed from inside inference/. The server runs on port 8173 and answers at POST /v1/systemone. That path is not an OpenAI-compatible route and does not pretend to be; it is the wire format the Decision Index evaluation kit speaks. The card's quickstart wires the kit to that endpoint directly, first for the text suite and then for the vision suite with an explicit pixel budget of 1,638,400 per image.

Two limits matter operationally and are stated in the card rather than discovered in production: four images per request, and 65,536 tokens of input for text-only or 10,000 tokens once images are attached. The processor configuration sets a maximum of 1,806,336 pixels per image, so a full-page screenshot will be downscaled before it is read.

The config file gives away its lineage

Open config.json and a set of keys appear that are not standard Q​wen3.5: jev_gdn_mode, jev_mask_version, jev_packing, jev_prefix_contract, jev_soft_choice. The prefix contract reads "all-input-before-first-answer; single-turn; no-thinking". The soft-choice block carries SHA-256 hashes of the Python files that implement the readout, pinned so an evaluator can confirm the code that produced a score is the code that shipped.

Read plainly, "j​ev" is the name of the shared harness this model was built into — the same family of harness that TypeSafe's J​ev models and the Decision Index kit are organised around. That is an inference from a string in a config file, not a statement from the vendor, and it is the most interesting unexplained thing in the release. It suggests Sing VL S1 is not a from-scratch experiment in typed decisions but a Q​wen3.5 vision checkpoint adapted into an existing protocol, with the protocol's internals shipped alongside so results can be reproduced.

There is a sibling in the same organisation that supports the reading. inclusionAI's LemonDecider, published in late October on a Q​wen3.5-4B base, is a text-only decision engine with a 255-option readout and a temperature calibration constant; its card describes the same choice/noul/score vocabulary and the same single-forward-pass design. Sing VL S1 looks like that line of work extended to images. The organisation has not said so.

What is missing is not cosmetic

• No licence. The model card's metadata declares tags for the architecture and the task but no licence field at all, and no LICENSE file sits in the repository. Without one, the default is not "open" — it is "all rights reserved, ask the author". A team that needs to ship this commercially has no permission to rely on yet, and should say so to its lawyers before it says so to its roadmap.

• No paper. The organisation has thirty-one papers on its Hugging Face profile and none of them is about this model. There is no method section, no training-data description, no statement of what the 4.54 billion parameters were tuned on.

• No announcement. The org card still points followers at its GitHub and its two X accounts for release news; neither channel has a post about this repository that could be found today. The org's GitHub has no SingVL repository.

• No hosted endpoint. No inference provider lists it. The only way to run it is to download 9.1 GB and serve it yourself.

• No independent result. Covered above, and worth repeating because it is the one that decides whether the 69.76 means anything to you.

How you would actually use it, if you decide to

The honest deployment story is a sidecar, not a replacement. Put the FastAPI server on a GPU box, point your existing pipeline at it for the specific sub-task where you currently parse a label out of generated text — intent routing, document-type triage, moderation flags, a yes/no gate before an expensive call — and leave the generative model where it is for everything that needs an explanation. The probability output is the point: you get a number to threshold, not a string to trust.

The thing that makes this cheap to try is that the API is one route with a flat JSON body. A wrapper function is a morning's work. The thing that makes it expensive to depend on is that a model with no licence, no paper and a one-day-old repository can change or disappear without notice, so pin the revision hash in the config and treat the endpoint as untrusted infrastructure behind a fallback.

That is also where routing earns its place rather than being a slogan. OrcaRouter already carries TypeSafe Jev 1.14 — typesafe/jev-1.14, at $0.042 per million input tokens — which means the decision-model pattern is callable today through one key alongside the generative models it would sit next to. Sing VL S1 is not on OrcaRouter and there is nothing to announce about it here; the relevant fact is narrower and more useful. You can measure the pattern on a working endpoint before you commit to serving a 9 GB checkpoint whose terms of use are blank, and if the checkpoint later earns its place, a second contract is not the cost of adopting it.

What would change this from a curiosity into a decision

Three things, in order of how much they would tell you. A licence file, because without it nothing else matters commercially. A submission to the Decision Index with a run directory attached, because the card's submission section describes exactly the artefact that would let a stranger check 69.76 and 73.94 rather than take them. And a second independent evaluation on a harness someone else controls, because a model that wins on a benchmark its own authors helped formalise needs a second opinion before it needs a launch post.

Until then the correct way to read this release is the way you would read a preprint with the methods section missing: the artefact is real, the numbers are the authors', and the interesting part is that a 4.5-billion-parameter vision model built for probability vectors rather than paragraphs now exists at all. Whether that becomes a category or a footnote depends on paperwork the repository does not yet contain.