SingVL_S1_4B vs WeVisOne VL 4B: The Quiet Release Against the Finished One

T​encent's WeChat Vision team published the paper, the weights, the inference package and the demo in October 2026 under Apache 2.0. That is what a finished release looks like. inclusionAI uploaded two safetensors shards on November 11, 2026, with no announcement, no paper, no licence and no hosted endpoint, and called it a day. WeVisOne VL 4B and SingVL_S1_4B are both roughly four and a half to five billion parameters, both built on Q​wen vision backbones, both concerned with getting something more precise than a paragraph out of an image — and they are doing opposite halves of that job. One draws masks over pixels. The other returns a probability over the answers you thought to offer it. Comparing them is less about which is better than about what a release is worth before it is finished.

The two capabilities, stated without flourish

WeVisOne-VL-4B — the name is a contraction of WeChat Vision All-in-One — unifies object-level perception and pixel-level perception in one model. Concretely, it does open-vocabulary object detection, referring expression localisation, object counting, pointing, OCR localisation, referring segmentation, and promptable alpha matting, which is the task of producing a soft continuous mask around something like hair or a thin branch where a hard boundary would look wrong. The mechanism is structured next-token prediction over dedicated coordinate tokens plus a 9.6-million-parameter pixel decoder shipped alongside the checkpoint as mask_decoder.pt. It returns None when a queried target is absent, which sounds trivial and is the single most useful behaviour in production grounding.

SingVL_S1_4B is a decision model. It takes a state — text, JSON, up to four images — and a dictionary of named questions, each with a declared type and candidate answers. Choice returns a probability per option, noul returns the probability that a yes/no statement holds, score returns a probability per level across two to ten ordered levels and the expected level. One forward pass, no generation, no coordinates. The repository includes a FastAPI server on port 8173 answering at /v1/systemone, and the model card is written entirely in that protocol's vocabulary.

So the split is clean. WeVisOne tells you where the thing is and what shape it has. SingVL_S1_4B tells you how likely it is that the thing is what you suspect. A pipeline that needs to cut a person out of a photograph has no use for a probability over three options, and a pipeline that needs to know whether an image passes a policy threshold has no use for a mask.

What the polished release put on its own card

WeVisOne-VL-4B's headline table compares it against the strongest open baseline in each benchmark, drawn from Qwen3-VL and Q​wen3.5 up to 9B, Rex-Omni, LocateAnything and task-specific segmentation and matting models. The rows that carry the argument:

• LVIS detection — F1 mean 55.3 against a 50.7 baseline.

• Dense200 dense detection — 71.3 against 58.7.

• RefCOCOg test referring — F1 mean 81.8 against 77.6.

• WeRef-Bench rejection rate — 96.4 against 73.0, which is the row that matters most for production grounding.

• PixMo-Count counting accuracy — 89.6 against 82.8.

• TotalText OCR — F1 mean 52.0 against 43.3.

• ReasonSeg referring segmentation — gIoU 71.2 against 65.0.

• P3M-500-NP box matting — MAD 7.3 against 15.4, lower being better.

• WeRefMat text matting — MAD 9.8 against 22.4.

Those are vendor figures on the vendor's evaluation pipeline, run at greedy decoding with at most 1,960 visual tokens per image, against baselines the vendor selected. The honest reading is that they establish the model works and roughly where it sits against the strongest open alternative on each task — not that the gaps are exact. The rejection-rate row is the most trustworthy in spirit, because 96.4 against 73.0 is a large enough difference that harness details are unlikely to explain it, and rejecting absent targets is the failure mode that ruins grounding systems in production.

There is also a 2B variant with its own column, which is useful: 78.2 on RefCOCOg test, 50.2 on rejection, 62.9 on the dense-detection row. The vendor published the smaller model's weaker numbers next to the larger model's stronger ones, which is the kind of disclosure that suggests the table was not assembled for a launch post.

What the quiet release put on its own card

SingVL_S1_4B publishes a "Public Decision Index 0.3" table: 69.76 on Text, 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 breakdown behind the vision aggregate runs from InfographicVQA at 96.31 and Winoground at 93.12 down to MMMU-Pro vision at 50.17 and R-Bench-M at 49.77.

Those figures have three problems that WeVisOne's do not. They are two days old. They are on a leaderboard whose scoring kit is public but whose private half is not, produced by a lab that has not submitted a run to the public reproduction kit — though the card does include instructions for filing one. And the repository carries no licence, which means the numbers are not the binding constraint on adoption; the paperwork is.

What the card does establish beyond argument is the interface: a per-question, per-option probability from a single forward pass, with the protocol's internals shipped as SHA-256 hashes in config.json so a run can be verified against the code that produced it. That is a real engineering decision and it is the most credible thing in the release.

The honest scoreboard has one shared row and it is empty

These two models have no benchmark in common. WeVisOne measures F1 at IoU thresholds, gIoU, and mean absolute difference on mattes. SingVL_S1_4B measures chance-corrected accuracy over option sets. There is no row where the same measurement was taken of both, and constructing one would require running both on the same data — which nobody has done, because one of the checkpoints is two days old.

What can be compared is the shape of the evidence. WeVisOne's release includes a technical report on arXiv, a project page, a GitHub repository with an evaluation pipeline, a cookbook, two demo spaces, and reproduction instructions that a third party can follow. SingVL_S1_4B's release includes a model card and an inference folder. One of those lets a stranger check the claims; the other asks the stranger to trust them.

Licensing, which decides more than either benchmark

WeVisOne-VL-4B is Apache 2.0. That covers T​encent-authored code and weights, with third-party components keeping their own terms, and it is why there are already GGUF quants, a ComfyUI node, a GGUF build for a Rockchip NPU and a Chinese-language integration package circulating within two weeks of release. Redistribution is permitted, so redistribution happened.

SingVL_S1_4B declares no licence in its metadata and ships no LICENSE file. inclusionAI has used explicit licences on its other releases — Apache 2.0 on LemonDecider, Q​wen's community licence on the SingProbe models — so the gap looks like an unfinished upload. But "looks like an oversight" is not a licence, and a compliance review cannot approve a model on the basis that its publisher probably meant to be permissive.

For anyone whose answer to "can we ship this" matters, the comparison ends here. WeVisOne can be shipped; SingVL_S1_4B cannot, yet. Everything else is a discussion about capability between two models where only one is currently available for the discussion to matter.

Which one fits which pipeline

• Cutting something out of an image — WeVisOne-VL-4B, and it is not close, because SingVL_S1_4B has no segmentation output at all.

• Counting objects in a crowded frame — WeVisOne-VL-4B. It reports 57.1 on the dense-detection row where the best baseline is 47.9, and counting is a first-class task rather than a prompted behaviour.

• Knowing a target is absent — WeVisOne-VL-4B, on the strength of the rejection row and the explicit None return. This is the capability that separates a demo from a system.

• Gating an expensive call on a confidence — SingVL_S1_4B's shape, if you can host it and the licence arrives. A probability per option is what a threshold wants.

• Classifying against a fixed taxonomy — SingVL_S1_4B in principle; its choice and score types map onto taxonomies directly, where a generative model has to be constrained back into one.

• Reading a page and getting both the fields and their positions — WeVisOne-VL-4B, which does OCR with localisation in the same pass as detection.

• Anything commercially deployed this month — WeVisOne-VL-4B, on licensing alone.

And a row that belongs to neither: if what you need is a written explanation of what is in the image, both models are the wrong tool. WeVisOne returns structured perception and SingVL_S1_4B returns probabilities; neither writes a paragraph.

What the comparison teaches about the market, not the models

The interesting thing here is not that one model is better. It is that two labs of comparable size released a 4-to-5-billion-parameter vision model in the same quarter and treated the release completely differently. T​encent shipped a paper, a decoder checkpoint with a stated parameter count, an evaluation pipeline, a licence and a project page, and the ecosystem built quants and integrations within a fortnight. inclusionAI shipped weights, an inference script and a benchmark table, and the ecosystem cannot touch any of it until the licence appears.

Both strategies are legitimate. A quiet drop lets a lab put a checkpoint out without a launch narrative, and the model card here is genuinely well written — the API examples run, the question types are documented with worked arithmetic, and the config ships hashes that make verification possible. But a checkpoint without a licence is a demonstration, not a product, and the difference shows up in exactly the place this comparison keeps returning to: whether anyone other than the author can build on it.

If you want to test the decision pattern before the paperwork arrives, the pattern itself is callable. OrcaRouter serves TypeSafe Jev 1.14 at $0.042 per million input tokens alongside 200-plus models under a single key, passing provider list prices through at 0% markup, with automatic failover across providers. Sing VL S1 is not on OrcaRouter and has no hosted home anywhere, so this is not an availability claim about it — it is the cheapest way to find out whether probability outputs beat masks on your own labels, before you spend a GPU on a checkpoint whose terms are blank.

WeVisOne-VL-4B you can simply download. That, in the end, is the comparison: one of these is a model you can use, and the other is a model you can read about.