ZwZ2-4B Shipped Without a Launch: What the Model Card Says and What It Won't

The ZwZ2-4B repository was created on 15 October 2026 at 06:11 UTC and, as of this morning, it has zero downloads and zero likes. That is not unusual for a checkpoint nobody has announced. What is unusual is that it also carries a complete model card: a method section, a training-data description, a quickstart block, and a headline result claiming 25 benchmarks and a 5.38-point improvement over Qwen3.5-4B, the model it was fine-tuned from. inclusionAI has published the documentation without publishing the announcement. The card is there; the press release is not; and the difference between those two things is the whole story of this checkpoint.

This piece is deliberately narrow about what it will assert. Everything below marked as a repository fact was read directly from the files — the configuration, the weights index, the dataset entry, the card itself. Everything marked as a vendor claim is inclusionAI's own number, reproduced as a claim and not as a result. There is no independent evaluation of ZwZ2-4B in existence, and anyone who tells you otherwise is reading the card back to you.

What actually landed, on what dates

The ZwZ2 family arrived in a window of about twenty-three and a half hours, in reverse size order. ZwZ2-4B came first, on 15 October at 06:11 UTC. ZwZ2-9B followed on 16 October at 03:42. ZwZ2-27B closed the set at 05:42 the same morning. The training dataset, inclusionAI/ZwZ2-RL-VQA, appeared three days later on 19 October. All three repositories, the dataset and the family collection carry timestamps from today, 27 October — the most recent maintenance pass on a family that still has no public announcement attached to it.

What did not happen: no entry on inclusionAI's blog and no release thread. The three checkpoints do share a collection, Zooming-without-Zooming-v2, four items strong, but a collection is a container rather than an announcement. The GitHub repository that hosts the first-generation code has not been updated at all — it still lists four models, ZwZ-2B, ZwZ-4B, ZwZ-7B and ZwZ-8B, all built on Qwen3-VL backbones. The v2 family is not in it. If you go looking for the code that produced ZwZ2-4B, you will find the code for its predecessor, and the README there will not mention the model you are holding.

The claim, stated as a claim

inclusionAI's card reports an overall accuracy of 49.41 for the base model and 54.79 for ZwZ2-4B, measured across 25 benchmarks grouped into six domains. The card says the model improved on every one of those benchmarks, and that the overall and domain averages weight each benchmark equally. The six domains it names are fine-grained perception, spatial reasoning, long-context understanding, multimodal agents and general visual reasoning, plus the fine-grained group that carries the headline.

Those two numbers are the entire quantitative case for the model. The per-benchmark scores exist on the page as figures — a PNG and an SVG — rather than as text, which means the eight representative results and the full twenty-five are visible but not machine-readable. The one benchmark set described in prose is MMLongBench*, whose visual core is listed as V-RAG, Image MM-NIAH and VH-Multi, over 25 non-Counting tasks. The card also notes that image resolution for "advanced models" is constrained by the serving platform, which is a fair caveat and also a reminder that a perception score is partly a statement about how the image reached the model.

A +5.38 average across 25 benchmarks is a large claim for a 4B-class model. It is also exactly the kind of claim that a vendor is best positioned to make and worst positioned to verify, because the vendor chose the benchmark set, the weighting and the resolution. Treat 54.79 as the number to reproduce, not the number to plan against.

The method: one pass, no zooming, no SFT

The training pipeline is called EvidenceConnect, and it is the second generation of an idea the lab published earlier this year under the name "Zooming without Zooming". The premise is that fine-grained visual perception usually gets solved at inference time — crop the region, re-ask the question, repeat — and that this is expensive and fragile. ZwZ2 instead teaches the model to attend to regional evidence inside a single forward pass.

The construction runs in four steps, as the card describes it. Related regions are grouped, three to seven per image, drawn from SA-1B. Questions are generated against privileged zoomed views of those regions, so the question-writer can see detail the policy model will never see at inference. Those questions are then filtered twice — by cross-model agreement, and by a drop-one-region test that checks whether removing a region actually changes the answer. Verified questions become four-option multiple-choice tasks, and the model is trained with GRPO. There is no supervised fine-tuning stage at all.

At inference the policy sees the full image with coloured region outlines drawn on it. The privileged crops exist only during data construction. Each model size in the family used roughly 20,000 training examples; the dataset card puts the exact figure at 20,662 multiple-choice questions, five downloads, and a creation date of 19 October.

The design has one visible consequence worth flagging for anyone who plans to evaluate it. A model trained only on four-option multiple choice is being optimised for a task shape that looks nothing like open-ended captioning or free-form visual QA. If your workload asks for a paragraph about an image, this checkpoint was not trained on that, and the card's limitations section says as much: hallucinated details, misread text, counting errors, spatial errors.

What the configuration proves, independent of the lab

There is a class of fact you do not have to take on trust, because it is arithmetic on published files. ZwZ2-4B weighs 4,539,265,536 parameters in BF16, stored as a single 9.08 GB safetensors file. The Hugging Face page rounds this to "5B", which is generous; the index says 4.54B. Its base, Qwen3.5-4B, weighs 4,659,865,088 — so the derivative is about 120.6 million parameters smaller than the model it was trained from.

That difference has a clean explanation in the config. Qwen3.5-4B declares one multi-token-prediction layer; ZwZ2-4B declares zero. The speculative-decoding head has been removed. Strip an MTP head from a 4.6B checkpoint and you land almost exactly on the observed gap. This is the same edit the 9B sibling carries, and it is the strongest available evidence that these are post-trained derivatives rather than freshly trained models.

The rest of the stack is unchanged from the base:

• Depth — 32 layers in both ZwZ2-4B and Qwen3.5-4B

• Width — hidden size 2,560, intermediate 9,216

• Attention — 16 query heads against 4 KV heads, head dimension 256, a 3:1 linear-to-full schedule (24 linear layers, 8 full)

• Vocabulary — 248,320 tokens, tied embeddings

• Context — 262,144 positions, M-RoPE in an 11/11/10 split, partial rotary factor 0.25

• Vision tower — 24 layers at hidden 1,024, patch size 16, 2×2 spatial merge

• Runtime — transformers 5.13.1, architecture identifier Qwen3_5ForConditionalGeneration

The chat template is worth one line because it changes how the model behaves out of the box. The released template runs in non-thinking mode and instructs the model to put its final answer inside <answer> tags. If you load this with a generic VLM harness that expects free text, you will get tags in your output and wonder why.

What the card does not tell you

Three gaps matter more than anything the card says.

The first is the licence. The repository has no LICENSE file and no licence tag. The card's own wording defers to "the licence accompanying the released weights and the applicable backbone licence" — but there is no accompanying licence file to defer to. The backbone, Qwen3.5-4B, is Apache-2.0, so the most likely reading is that the derivative inherits it. Likely is not the same as declared, and a legal review will not accept a card sentence that points at a file which does not exist.

The second is the paper. The v2 method — EvidenceConnect, the drop-one-region filter, the GRPO-only training — has no public write-up. The card's paper link is still the literal placeholder PAPER_URL. The only arXiv identifier attached to the repository, 2602.11858, is the first-generation paper, which describes region-to-image distillation on Qwen3-VL backbones: a different method, a different base family, and a different set of numbers. The dataset card inherits the same identifier, so the tag propagates the mismatch.

The third is the absence of any independent measurement. No leaderboard has scored ZwZ2-4B, no replication exists, and the download counter has not moved off zero. A model can be excellent and unmeasured at the same time; the point is only that nothing outside the lab has yet confirmed the 54.79.

The naming trap you will walk into

Search for this model and you will find a different one first. The first-generation ZwZ-4B is a real, documented checkpoint built on Qwen3-VL-4B, and it publishes an average of 66.86 against its backbone's 61.52, with a ZoomBench score of 55.74. Those numbers are widely quoted and they are not ZwZ2-4B's numbers. They belong to a different base model, a different training recipe and a different benchmark suite.

The two names differ by a single character. A comparison table that mixes them will look internally consistent and be wrong. If a vendor page or an aggregator shows ZwZ2-4B at 66.86, that is the error you are looking at.

If you need a small vision model this week

ZwZ2-4B is not in the OrcaRouter catalogue, and this piece is not going to pretend otherwise. It is a checkpoint you download and serve yourself; there is no hosted endpoint for it anywhere that the Hugging Face page acknowledges, and the page explicitly reports that no inference provider serves it.

What the catalogue does have is the practical alternative for the same job: qwen/qwen3-vl-8b-instruct at $0.18 per million input tokens and $0.70 per million output, with a 131K context, served through the same OpenAI-compatible endpoint as everything else. If your plan was to evaluate ZwZ2-4B against a hosted vision baseline, that is the baseline, and running both from one key means one client library and one bill instead of two integrations. OrcaRouter routes 200-plus models with automatic failover, so a vision path built on that endpoint does not depend on a single upstream staying healthy.

For a model with zero downloads and no licence file, that is the honest recommendation: rent the comparison, own the experiment.

What would change this assessment

Three events, in descending order of importance. A licence file, which would turn a legal unknown into a legal answer. A public v2 paper, which would let someone check the EvidenceConnect filtering for leakage — the drop-one-region test is a good idea, and good ideas are exactly what need auditing. And any third-party score at all, because a single reproducible benchmark on a neutral harness would do more for this checkpoint than another thousand words of card.

Until one of those lands, ZwZ2-4B is a well-documented, unannounced, unmeasured 4.54B perception specialist whose own card says it beats its base model by 5.38 points. That is a genuine and specific claim. It is not yet a result, and the distinction is the only thing worth carrying out of this repository.