ZoomGain-9B Shipped Without an Announcement: What the Repository Actually Says

The paper went up on arXiv on 10 November 2026. The GitHub repository was created a day earlier, on 9 November, and its first commit is titled "Release ZoomGain init". The weights appeared on Hugging Face on 10 November at 11:42 UTC, followed by a GGUF repository on 11 November at 03:16. There is no launch blog post, no vendor announcement, no hosted endpoint, and no licence tag on the checkpoint repository itself. That is the entire public record for ZoomGain-9B, a 9.4-billion-parameter vision-language model from inclusionAI, the open-model arm of Ant Group, and the reason this piece is written as a repository reading rather than a launch story.

What makes it worth the attention is not the stealth. It is that the repository is unusually complete for something nobody announced: the paper, the 1,000-question benchmark, the 9,000-record reinforcement-learning dataset, the evaluator, the training launcher, two checkpoints at 4B and 9B, and three GGUF quantisations all landed inside roughly two days. You can read the method, re-run the evaluation and reproduce the training today. The only thing missing is someone from the company saying it exists.

What actually landed, in order

• 9 November 2026 — the inclusionAI GitHub repository is created. It carries the Apache 2.0 licence and the data-construction and training code, not the weights.

• 10 November 2026 — arXiv:2611.14076v1, "ZoomGain: Mining Global-to-Zoom Perception Gaps for Fine-Grained Visual Supervision", by Haozhan Shen, Liangbo He, Jun Lan, Huijia Zhu, Weiqiang Wang, Tiancheng Zhao and Jianwei Yin, with affiliations at Ant Group, Zhejiang University, Om AI Research and the Binjiang Institute.

• 10 November 2026, 11:42 UTC — inclusionAI/ZoomGain-9B and inclusionAI/ZoomGain-4B are created on the Hub, each with a single 18.82 GB and 9.08 GB BF16 safetensors file respectively.

• 11 November 2026, 03:16 UTC — inclusionAI/ZoomGain-9B-GGUF appears with Q4_K_M, Q8_0 and F16 builds plus matching mmproj projector files, which is the step that turns a research checkpoint into something a laptop can attempt.

A small detail worth noting because it is exactly the kind of thing a coordinated launch does not produce: the model card's own "Updates" section dates the paper, the benchmark, the dataset and the checkpoints all as 2026.11.11, while the repository's creation timestamp and the arXiv submission date both say 10 November. The card was written after the fact, in one pass, by someone tidying up.

The idea, in one paragraph, because it is genuinely one idea

A vision model can often answer a question about a region of an image that it fails when the same region appears inside the whole frame. The authors call this the global-to-zoom perception gap, and they turn it into supervision. Stage one: several models independently locate a target described in the full scene, and their agreement identifies a region to crop. Stage two: captions of the crop and captions of the full scene are compared, and any detail that the crop reveals but the full view does not becomes a question asked about the original image. Stage three: answers are re-checked on the crop, ambiguous questions and questions answerable from language priors are filtered out, and plausible distractors are generated for the multiple-choice format.

Training is GRPO on 9,000 records built over 7,713 source images — one epoch, batch size 64, eight rollouts per question — using the bundled EasyR1 code. Evaluation defaults to temperature 0 with Q​wen3.5 thinking disabled, and OCR and counting questions stay open-ended while attribute questions become multiple choice. None of that is unusual in itself; what is unusual is that the pipeline that generated the training data also generated the benchmark, and the paper is explicit about it.

Two numbers that are not the same number

On ZoomGainBench — the 1,000-question, 941-image, English-and-Chinese benchmark built by this same pipeline — ZoomGain-9B scores 55.48. Its base model, Qwen3.5-9B, scores 32.13. That is a 23.35-point jump on a benchmark the model's own training data was mined from, and the paper labels the column "Held-in" for precisely that reason.

The number to carry into a decision is the average over thirteen benchmarks the training never touched, which excludes ZoomGainBench: 66.10 for ZoomGain-9B against 59.75 for Qwen3.5-9B. A 6.35-point average gain from 9,000 samples is a solid, believable result. Both numbers are real. They answer different questions, and any page that quotes 55.48 without saying "held-in" is telling you half of it.

There is a second oddity in the same table that nobody seems to have noticed yet: the smaller sibling beats the 9B on the held-in benchmark. ZoomGain-4B scores 55.81 on ZoomGainBench against the 9B's 55.48, while losing the held-out average 63.96 to 66.10. On the benchmark built from the training distribution the two are level within noise; once you leave that distribution, size reasserts itself. If you were choosing between them on the strength of a single headline number, you would pick the wrong one.

What 9,000 samples bought, benchmark by benchmark

All of the following are author-reported figures from Table 3 of the paper, run by the authors' own harness on the released evaluator, and not independently reproduced by anyone:

• HR-Bench 4K — 90.25, up from 82.13. This is the number the abstract leads with, and it is the strongest claim in the release.

• V*Bench — 93.19, up from 86.91. Small-object detection at high resolution.

• VisualProbe — 56.31, up from 44.66. The largest relative gain outside the held-in benchmark.

• O3-Bench — 68.12, up from 56.52. Multi-step perception and reasoning.

• MMVP — 70.00, up from 60.00. The classic test of paired images that differ in one detail.

• MMStar — 73.87, up from 71.73. General multimodal competence, and the smallest gain of the set.

• BabyVision — 17.53, up from 15.98. Still near the floor, and a useful reminder that the gains are concentrated in the fine-detail family rather than spread evenly.

Read together, the pattern is consistent with the method rather than with a general capability bump: the biggest deltas sit on benchmarks that reward finding a small thing in a large frame, and the smallest sit on benchmarks that reward broad world knowledge and ordinary visual reasoning. That is what a targeted RL run on 9,000 cropped-detail questions should look like.

What is knowable, what is claimed, and what is simply not confirmed

Knowable from the repository, today, without taking anyone's word for it: the architecture, because the config is public and identifies a Qwen3.5-9B text trunk at 32 layers, hidden size 4,096, intermediate 12,288 and 262,144 maximum positions, wrapped in a vision tower; the training data, because the 9,000-record Parquet and the source images ship as a dataset; the method, because the prompts are in the paper's appendix; and the evaluation, because the evaluator, the annotations and the images are all downloadable.

Claimed but not confirmed: every benchmark score above. The paper reports them on the released harness, which makes them more checkable than most vendor tables — you can re-run them — but nobody outside the author list has done so yet, and no independent lab has published a ZoomGain-9B result.

Not confirmed at all, and worth being blunt about:

• The checkpoint licence. The code is Apache 2.0 and the card says so. The card then defers the model and dataset terms to "the Hugging Face cards for the terms of each dataset and checkpoint", and the checkpoint repository carries no licence tag in its Hub metadata. That is a gap to close by reading the files before you ship, not a promise you can rely on.

• The context length. The card states none. The config's 262,144 maximum positions are inherited from the Qwen3.5-9B trunk, which is not the same as a tested property of the post-trained checkpoint, and every evaluation in the paper is a single image with thinking disabled.

• General assistant behaviour. Nothing in the release tests chat, tool calling, long-form generation or multi-turn use. The model was trained to answer questions about photographs, and that is the only capability with evidence behind it.

• Any host. The Hub's provider mapping for the repository is empty. There is no vendor API and no third-party endpoint listed, so "try it" currently means "download 18.82 GB".

Running it this week, if you want to

The GGUF path is the cheap one: a 5.63 GB Q4_K_M build plus a 0.52 GB mmproj projector, or 9.53 GB at Q8_0, or 17.92 GB at F16. Image input needs both files — the projector is not optional, and a llama.cpp build with Q​wen3.5 multimodal support is the prerequisite. The reference path is heavier but better documented: vLLM 0.19.1 with transformers 5.7.0, exactly the versions the authors evaluated with, on a BF16 checkpoint that wants a 24 GB card at minimum and comfortable headroom above that.

Worth budgeting for the evaluation protocol as well as the weights. The authors run at temperature 0 with thinking disabled and score attribute questions by exact match on the chosen option, counting by exact match on the predicted integer, and OCR by ANLS with a 0.7 floor. If you evaluate with thinking on, or with a different prompt template, you are measuring something the authors did not report and should not expect their numbers to transfer.

OrcaRouter does not host ZoomGain-9B, and this piece is not going to pretend otherwise. What we do have is one key across 200-plus models at provider list price with no markup added, which matters here for a specific and practical reason: the fastest way to find out whether your pipeline's errors are fine-detail errors is to run the same images through a hosted vision model and compare failure cases. Qwen3-VL-8B-Instruct on our catalogue is $0.18 per million input tokens and $0.70 per million output at a 131K context, takes images and video, and will not match a model trained on crop-mined questions — but it will tell you within an afternoon whether the GPU is worth buying.

What would change this story

Four things, in ascending order of how much they would matter. An announcement from inclusionAI, which would turn a repository reading into a launch and probably arrive with a licence clarification. A checkpoint licence tag, which would settle the one question that currently blocks commercial adoption. A hosted endpoint, which would remove the 18.82 GB barrier for everyone who does not own the hardware. And an independent evaluation, which is the one that would actually move the needle: the claim that a 9B model reaches 90.25 on HR-Bench 4K is the kind of number that either survives a second lab or quietly disappears.

Until then, the honest summary is that ZoomGain-9B is a well-documented research release with a legible method, a reproducible harness, an unusually honest held-in/held-out split, and no third party standing behind any of it. That is more than most quiet drops offer, and less than a product.