SingVL_S1_4B vs Moondream 3: Coordinates or Probabilities

Both of these models promise structure instead of prose, and that is where the resemblance ends. Moondream 3 answers in coordinates: ask it to find the people in a photograph and it returns boxes with labels, ask it where the handle is and it returns points, ask it to caption and it returns a sentence. SingVL_S1_4B answers in probabilities: hand it options and it returns a number per option, hand it a statement and it returns the chance the statement is true. One tells you where something is. The other tells you how sure it is. Choosing between them is a question about which of those two answers your pipeline was always going to need, and the two rarely substitute for each other.

The two shapes of "structured output"

Moondream 3 is a mixture-of-experts vision model — 9B total parameters, 2B active — that the vendor, Moondream, describes as built for visual reasoning and detection at deployment-friendly cost. Four skills are native rather than prompted: query for visual question answering, detect for open-vocabulary boxes, point for centre points, and caption. All four return structured output rather than conversational text, and the weights are free to download and run locally through the vendor's own inference engine, Photon, on NVIDIA GPUs from Ampere onward or on Apple silicon. The model appeared as a preview in September 2025 and was superseded by Moondream 3.1 in July 2026, which is the current generation of the same architecture.

SingVL_S1_4B is a decision model, uploaded by inclusionAI on November 11, 2026 with no announcement. It takes a state — text, JSON or up to four images — and a dictionary of named questions, each declaring a type and its candidate answers. Choice returns a probability per option, noul returns the probability that a yes/no proposition holds, score returns a probability per level across two to ten ordered levels plus an expected level. One forward pass per request, no token generation, no free-form output of any kind.

Put concretely: give Moondream 3 a warehouse photo and the word "pallet", and it hands you boxes. Give SingVL_S1_4B the same photo with the question "is this frame safe to dispatch" and candidates yes and no, and it hands you 0.91 and 0.09. Both are structured. Neither can produce the other's output.

What the 9B model is actually for

Moondream's own release material is unusually candid about the limits of benchmarks, and then publishes a dense set anyway. On detection at [email protected] the numbers that matter are COCO at 81.46, ODinW-13 at 93.94, Dense200 at 74.61 and SKU-110K at 52.77 — the last two being crowded-scene sets where a general vision model typically collapses. Referring detection lands at 87.1 on RefCOCO-M. Against those, LVIS at 67.4 and HumanRef at 70.39 are the rows where it loses to a competitor, which the vendor prints rather than hides.

That profile — strong on dense and open-vocabulary detection, weaker on fine-grained referring where a specialist wins — is what the 2B-active MoE buys. The model is fast because most of its parameters are not used per token, and it is accurate at counting and localising because those were trained as first-class skills rather than emerging from captioning.

The commercial shape is worth noting because it is the opposite of a research drop. The weights are free under the Moondream Model License, running them locally requires no API key, and the hosted service is priced per token: $0.30 per million input and $1.00 per million output for real-time, halved to $0.15 and $0.50 for batch. There is a free tier, a playground, a documented local path and a licence text you can read. Nothing about the release is ambiguous.

What the 4.5B model is actually for

SingVL_S1_4B's card publishes a "Public Decision Index 0.3" table: 69.76 on the Text panel and 73.94 on Vision, against TypeSafe Jev 1.14 at 60.70 text, StartLux-Decision v1.1 4B at 67.8 and 67.90, and Perplexity Decider Flash v1.2 at 60.1 and 72.5. The per-dataset rows show where the aggregate comes from — InfographicVQA 96.31, Winoground 93.12, key-information extraction 95.82 at the top; MMMU-Pro vision 50.17, R-Bench-M 49.77 and hateful-meme moderation 71.64 at the bottom.

Those are the vendor's own figures on a leaderboard its own authors helped formalise, produced by scoring supplied options rather than generating answers, and nobody has reproduced them. The repository was two days old when this was written, has zero downloads, ships no licence file, and no paper exists. The Decision Index's public reproduction kit shows no submission from this model.

So the usable claim is narrow and structural rather than numeric: a single forward pass that returns a calibrated probability per candidate is a different primitive from a model that returns coordinates or text. It is the primitive you want when the downstream decision is a threshold — when somebody has already decided that above 0.8 you act and below 0.3 you escalate. It is not the primitive you want when the downstream step needs to know where.

The overlap is smaller than a shared dataset name suggests

Both cards touch visual grounding, and it would be easy to line up numbers and declare a winner. Don't. Three mismatches make it meaningless.

• Task shape — Moondream's detection scores come from emitting boxes and matching them to ground truth at an IoU threshold. SingVL_S1_4B's vision scores come from choosing among options someone else wrote. A model that ranks candidates correctly is not thereby a model that can localise.

• Harness — Moondream reports [email protected] on named detection sets under its own evaluation settings. SingVL_S1_4B reports accuracy inside the Decision Index's chance-corrected, coverage-adjusted suite. Same word, different measurement.

• Authorship — both sets are vendor-reported. Moondream's has had over a year of third-party scrutiny and community ports; SingVL_S1_4B's has had none. One of those numbers has been argued with in public.

Where the two genuinely could be compared is a task both can express: does this image contain a violation, and which category. Moondream would answer with a detection or a query response; SingVL_S1_4B with a choice distribution. That is a real experiment somebody should run, and it has not been run.

A worked shape, because the API explains the product

The clearest way to see the difference is to write the request each model would receive for the same job — triaging a delivery photo before it goes to a claims adjuster.

With Moondream 3, the call is detect(image, "damaged parcel"), and the response is a list of objects with bounding boxes and confidence, or an empty list if nothing matched. You learn whether damage was found and where. Then caption(image) gives you a sentence describing the scene. Two calls, two structures, both about content.

With SingVL_S1_4B, the call carries the questions in the request. One choice question asks whether the parcel is intact, damaged or missing, with three candidates. A noul question asks whether the label is legible. A score question rates image quality across four ordered levels. The response carries a probability per candidate for each question and an expected level for the score. One call, three answers, all about confidence.

The Moondream path gives a claims system something to display. The SingVL path gives it something to branch on. A team building the first will find the second useless; a team building the second will find the first requires a parser it does not want to write. This is the whole decision, and it is settled before either vendor's benchmark is opened.

Running them is not comparable either

Moondream 3.1 runs on a workstation GPU or an Apple laptop through Photon, with weights downloaded from the Hub on first use and no key required for the base model. A hosted endpoint exists with published per-token prices, a free tier and a documented quickstart, and the vendor publishes a local-inference guide covering supported hardware, streaming, finetunes and serving through Triton. Every path is documented and every price is public.

SingVL_S1_4B runs on a GPU you provide, from a 9.1 GB BF16 download, via a Python 3.12 environment and a FastAPI server bound to port 8173 that answers at POST /v1/systemone. The instructions are complete and the curl examples work, which is more than most quiet drops manage. What does not exist is any hosted option, any quantised build, any framework integration, or any licence permitting you to redistribute what you downloaded.

That last absence has a commercial edge. Moondream's licence is a custom one — not open source in the OSI sense — and the vendor is explicit that hosting the weights as a cloud service requires its licence, which is why the hosted tier exists and is priced. SingVL_S1_4B has no terms at all, which is not the same as permissive: with no grant, nobody outside inclusionAI has permission to use it commercially, and a legal review will stop there.

Where the decision lands

If the job is finding things — counting stock, locating a component, reading a layout into boxes, pointing at a handle — Moondream 3 is the only one of the two that can do it, and its crowd-scene numbers are the reason to prefer it over a general vision model for that work. It also has a price sheet, a licence, a local path and a year of other people using it.

If the job is deciding — a gate, a triage, a threshold over a fixed taxonomy where a number is more useful than a sentence — SingVL_S1_4B's design is the better fit, and its output removes an entire class of downstream parsing. But "better fit" is a statement about architecture on day two of a release with no licence, and it should be read as a reason to watch rather than a reason to build.

A team that wants to test the decision pattern now, rather than wait for inclusionAI to finish the paperwork, can call it today: OrcaRouter serves TypeSafe Jev 1.14 — typesafe/jev-1.14, $0.042 per million input tokens — on the same key as 200-plus other models, at provider list price with 0% markup, so a price change from any vendor lands the same day. Sing VL S1 is not on OrcaRouter and has no hosted home anywhere, so this is not an availability claim; it is the cheaper way to find out whether probability outputs beat boxes on your own data. If they do, revisit the checkpoint when it has terms. If they don't, you have saved a GPU and a legal review.

The two models will keep being mentioned in the same breath because both avoid the chat interface. That is a surface similarity. Underneath, one is a locator and one is a scorer, and the pipeline you are building already knows which of those it needs.