WeVisOne-VL-2B vs Moondream 3: A Licence and a Deadline

On 13 November 2026 the hosted endpoint that serves Moondream 3 Preview stops answering. That date, not any benchmark, is what makes this comparison urgent: teams that built on Moondream's cloud have roughly three weeks from the shutdown notice to decide what replaces it, and the weights they would fall back on carry a licence most of them have never read closely. WeVisOne-VL-2B is the obvious thing to put beside it — a 2B perception model from Tencent's WeChat Vision team, published on 22 October 2026 under Apache 2.0, doing the same grounded work: boxes, points, counts and masks.

The two models are the closest pair in this series. Both are built around pointing at things rather than describing them. Both output masks. Both were trained to say "not there" when the target is absent. And yet almost nothing either vendor published can be subtracted from the other, because they measured different things on different harnesses at different times. What can be compared is the licence, the output format and the deployment path — and on those three, the answer is not close.

Why these two are in the same conversation

Moondream 3 Preview, released 18 September 2025, made the case that a small model could do grounded perception natively: detect, point, count, segment and answer, all from one checkpoint, without orchestrating a separate detector and a separate reasoning model. It is a sparse mixture of experts at 9 billion total parameters with 2 billion active — 24 layers, the first four dense, the rest carrying 64 experts with 8 activated per token — a 32,768-token context, a custom SuperBPE tokenizer and a SigLIP-based vision encoder.

WeVisOne-VL-2B arrives a year later making a narrower version of the same argument. It is 2B parameters built on Qwen3-VL, and it splits the work in two: the language model emits coordinates through dedicated tokens, and a separate 9.5-million-parameter decoder turns each marked instance into a mask or a continuous alpha matte. Tencent did not try to make a general assistant. It tried to make a perception engine, and said so in the limitations section.

The family resemblance is real, which is why the differences matter more than they would between unrelated models.

The capability maps, side by side

• Parameters — 9B total with 2B active for Moondream 3 Preview vs about 2B dense for WeVisOne-VL-2B, plus a required 9.5M-parameter decoder.

• Mask output — SVG path strings plus a bounding box on Moondream, which its docs describe as native vector masks; raster alpha mattes at stride-1 full resolution on WeVisOne, including soft transparency for hair and fur.

• Reasoning — Moondream 3 has an optional chain-of-thought mode and reasons about an image by pointing into it; WeVisOne generates structured output with no chain-of-thought at all, a choice Tencent made for inference efficiency.

• Context — 32,768 tokens published on the Moondream card vs no published context length for WeVisOne, whose paper describes training at 4,096 and 8,192 tokens and inference at up to 1,960 visual tokens per image.

• Input — images on both, but Moondream ships a documented cloud API with streaming partial results; WeVisOne rejects video grids in its pixel decoder and is single-turn by design.

• Licence — Business Source License 1.1 with an Additional Use Grant on Moondream, Apache 2.0 on WeVisOne.

• Release — 18 September 2025 vs 22 October 2026, thirteen months apart.

• Hosting — Moondream Cloud until 13 November 2026, then self-hosted only; no hosted endpoint for WeVisOne from Tencent or from any provider we can verify.

The numbers, with their instruments attached

This is where the comparison becomes an exercise in reading footnotes. Every figure below is vendor-reported; neither model has an independent evaluation on Artificial Analysis or anywhere else we can find.

• RefCOCO — Moondream's launch chart lists 91.1, alongside 88.6 on RefCOCOg and 81.8 on RefCOCO+. Those are detection-style accuracy figures, and Moondream itself described the chart as not a fair fight: the frontier baselines were queried through templates because they have no native grounding, and several preview scores came from 100-question samples rather than full evaluation sets.

• RefCOCO on WeVisOne — Tencent does not publish a RefCOCO detection number. Its referring detection table reports RefCOCOg at 77.1 mean F1 for the 2B, and its segmentation table reports RefCOCO validation at 79.7 cIoU. Different task, different statistic, different split.

• Segmentation — Moondream's March 2026 segmenting update reports mIoU of 83.2 on RefCOCO validation, 79.1 on RefCOCO+ and 80.7 on RefCOCOg, with a separate 88.2 on the high-fidelity RefCOCO-M. WeVisOne reports 79.7, 73.1 and 74.4 on the same three splits — but as cIoU, which the paper states explicitly. Mean IoU and cumulative IoU weight images differently, so the gap here is suggestive rather than measured.

• Counting — Moondream reports CountBenchQA at 93.2. WeVisOne reports CountQA at 47.9 exact-match, PixMo-Count at 90.9 and 88.1, and CLOC at 27.5 and 26.9. CountBenchQA and CountQA are not the same dataset, and exact-match on a counting set is unforgiving in a way accuracy on a multiple-choice set is not. The spread inside WeVisOne's own counting row — 90.9 on one dataset, 27.5 on another — is a better warning about counting benchmarks than any cross-vendor comparison.

• Detection — Moondream lists COCO detection at 51.2. WeVisOne lists COCO mean F1 at 55.2 across IoU thresholds. Whether those are the same statistic is not stated on either side, which is exactly the problem.

What emerges is not a winner. It is two models that both claim strong grounded perception, measured with instruments that were never calibrated against each other, one of them published on a chart its own author described as unfair.

The licence is the real fork in the road

Moondream 3 Preview's weights are released under the Business Source License 1.1 with an Additional Use Grant whose short name is "No Third-Party Service." The grant covers personal projects, research, benchmarks, fine-tunes, merges, quantizations and weight deltas. What it does not cover is standing up a service that lets third parties use the model — the thing most companies actually want to build.

That is a material constraint, and it is the one place where the comparison has a clean answer. WeVisOne-VL-2B's card grants Apache 2.0 for Tencent-authored code, with third-party components keeping their own terms. No revenue threshold, no field-of-use carve-out, no conversation with a lawyer about whether your product counts as a third-party service.

If you are choosing between these two to build something commercial, the licence decides more of the outcome than any benchmark in the previous section. A model you cannot legally serve is not a cheaper option; it is a different product.

If you are on Moondream Cloud right now

The shutdown notice went up alongside the reason for it. Moondream and its successor Moondream 3.1 are now part of Prime Intellect, which announced the acquisition on 13 October 2026, with founders Vik Korrapati and Jay Allen joining to lead inference. Moondream's own post is dated 14 October. The notice was one line: on 13 November 2026, thirty days out, Moondream Cloud shuts down, with unspent purchased credits refunded.

Two things are worth separating. The models are not being withdrawn — the weights stay where they are and the licence does not change. What is ending is the hosted path, the one that most teams actually call. Photon, Moondream's inference engine, is part of what Prime Intellect bought, and its pitch is throughput: 34 requests per second for Moondream 3.1, more than 60 inferences per second on a single H100. Those are Prime's numbers and Prime is now the party selling them.

For anyone migrating, the honest options are three. Self-host the Moondream weights you already have and accept the BSL terms. Move to a hosted model that is not going away, accepting that most of them do not output masks. Or move to an Apache-2.0 open-weights model and own the deployment — which is where WeVisOne-VL-2B enters, with the caveats that it is not a drop-in replacement and that its evidence is thinner.

It is not a drop-in for three concrete reasons. The output format differs: SVG paths versus raster alpha mattes, which changes your downstream compositing code. There is no chain-of-thought mode, so any pipeline that relied on Moondream's grounded reasoning traces loses them. And there is no published context length, so a workflow that fed Moondream long multi-image prompts has no stated ceiling to design against.

This is also where a routing layer earns its place, and it is worth being precise about what it can and cannot do. OrcaRouter does not serve Moondream 3 or WeVisOne-VL-2B; neither is on our catalogue, and we will not pretend a migration is a URL change. What we can do is keep the hosted side of your stack portable: one API key across more than 200 models at provider list price with no markup, automatic failover when a provider degrades, and a routing DSL that lets you move traffic between hosted models without touching application code. If the lesson of 13 November is that a single hosted dependency is a single point of failure, the fix is the abstraction layer, not a different single dependency.

What to decide before 13 November

If you are running Moondream 3 Preview through the cloud API and nothing else, the decision is not really Moondream versus WeVisOne. It is whether you want to own inference at all. Self-hosting Moondream keeps the behaviour you have already tuned against, under a licence that permits internal and research use but not resale as a service. Self-hosting WeVisOne gives you cleaner terms and soft alpha mattes, at the cost of a model nobody independent has tested and a decoder that only works on single images.

If you are starting fresh — no existing pipeline, no tuned prompts, no sunk integration — the calculation is simpler. WeVisOne-VL-2B is Apache 2.0 and 2B, Moondream 3 Preview is source-available and 9B with 2B active, and the only benchmark both publish in a form that invites comparison is segmentation, where the metric definitions quietly diverge. Choose on the licence and the output format, then measure on your own images, because neither vendor has given you a number you can trust across the gap.