Title card for the ZoomGain-4B article: the headline ZoomGain-4B, a subtitle reading inclusionAI published the weights on 11 November 2026 and announced nothing, and three chips reading 4.54B parameters, Vision-language and No hosted API.
Engineering & Research

ZoomGain-4B Shipped on November 11 With No Announcement Behind It

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

The first commit in inclusionAI/ZoomGain is stamped 2026-11-09 at 02:13 UTC and its message is "Release ZoomGain init." The weights for ZoomGain-4B appeared on Hugging Face on 2026-11-10 at 11:38 UTC. The arXiv listing, the evaluation set and the reinforcement-learning corpus that produced it are all dated 2026-11-11, the same day the model card was last touched. What has not appeared, anywhere I could find, is a launch post. There is no news entry on inclusionAI's site, no announcement on the organisation's Hugging Face profile beyond the repository itself, no hosted endpoint, and no coverage from anyone. ZoomGain-4B is a vision-language model built by reinforcement learning on top of Qwen3.5-4B, and as of this morning the only people who know it exists are the ones who happened to look at a repository listing. This is a what-the-repo-says article: every figure below is either a timestamp you can verify yourself or a number the authors put on their own card, and the places where that is not enough are marked as such.

For a reader who wants the shape of it before the detail: the model is a 4.54-billion-parameter multimodal checkpoint that answers multiple-choice questions about fine visual detail — small text in a photograph, an object count in a cluttered scene, an attribute that is only legible when you crop in. The authors' claim is that a vision model looking at a full image misses details it can read perfectly well once the same region is zoomed, and that this gap can be turned into training signal. They then built a 1,000-question benchmark out of the same machinery, and on that benchmark the trained 4B scores 55.81 against 30.03 for the model it was trained from. Those two numbers are the whole case, and they come with a caveat that the paper states itself: the benchmark is built by the same pipeline as the training data, which the paper calls "held-in".

What landed, and what the timestamps say

The release is unusually complete for something nobody announced. Six artefacts, all within a 48-hour window:

• The paper. "ZoomGain: Mining Global-to-Zoom Perception Gaps for Fine-Grained Visual Supervision," arXiv 2611.14076, submitted 10 November 2026, by Haozhan Shen, Liangbo He, Jun Lan, Huijia Zhu, Weiqiang Wang, Tiancheng Zhao and Jianwei Yin. Seven authors, cs.CV and cs.AI.

• The 4B checkpoint. inclusionAI/ZoomGain-4B, safetensors in BF16. The file metadata counts 4,539,265,536 parameters. Apache 2.0 on the code; the card defers checkpoint and dataset terms to their own repositories.

• The 9B checkpoint. inclusionAI/ZoomGain-9B, 9,409,813,744 parameters, created four minutes after its smaller sibling.

• Quantised builds. ZoomGain-4B-GGUF landed on 2026-11-11 at 03:09 UTC with three precisions plus matching vision projectors: Q4_K_M at 2.71 GB, Q8_0 at 4.48 GB, F16 at 8.42 GB, each with an mmproj file for the image encoder. The 9B GGUF followed.

• The benchmark. inclusionAI/ZoomGainBench, 1,000 questions over 941 images, English and Chinese text for every item, plus a reference answer, a scoring rule and an evidence bounding box in pixel coordinates.

• The training data. inclusionAI/ZoomGain-RL-Data, 9,000 records and 7,713 source images in Parquet, with the reward metadata the training script expects.

The quietness is not a mystery and it is not sinister. It is what a research group's internal release looks like when the paper clears arXiv before anyone writes the blog post. The tells are all there: three commits on the morning of 11 November are all titled "update readme" fourteen minutes apart, the GitHub repository has three stars and no description field filled in, both checkpoint repositories show zero downloads and zero likes, and the 4B model page carries a warning about empty or missing YAML metadata. None of that is evidence about the model. It is evidence about the rollout.

Screenshot of the inclusionAI/ZoomGain-4B model card on Hugging Face, captured 11 November 2026, showing the repository header, the parameter metadata and the opening of the model card text.

The method, in three moves

The interesting part of ZoomGain is not the model, it is how the questions were made. The pipeline runs in three stages and only the first one is conventional.

• Ground the target. Several models independently locate a region described in the full scene. Where they agree, that agreement is taken as the crop worth keeping. Using disagreement as a filter means no single model's blind spot defines the target set.

• Mine evidence. Captions of the same region are compared at two zoom levels. Whatever the crop reveals and the full image does not is turned into a question about the full image. This is the mechanism that produces the "global-to-zoom" gap the title refers to: the question is answerable, but only if the model resolves something at the edge of its perception.

• Verify and distract. Answers are cross-checked against the crop, ambiguous questions are dropped, questions that can be answered without the image at all are dropped, and multiple-choice distractors are generated. The ablation says this last step is doing real work — removing dedicated distractors costs 8.96 points on the held-in benchmark, the largest single drop the paper reports.

Training itself is not exotic. GRPO through the bundled EasyR1 framework, one epoch, batch size 64, eight rollouts per question, which works out to 140 steps on 9,000 records. The repository ships a smoke test that runs a single step including rollout, reward and checkpoint save, and full launchers for both sizes. PyTorch 2.10, FlashAttention 2.8.3, vLLM 0.19.1 and transformers 5.7.0 — all pinned, which is the difference between a repository that reproduces and one that merely describes.

The number that matters, and the caveat that shrinks it

ZoomGainBench is deliberately hard. The strongest model on it is GPT-5.6-Sol at 61.08 overall, and that is 61.08 out of 100 on multiple-choice questions about photographs. Claude-Opus-5 sits at 59.86, Gemini-3.6-Flash at 59.41. The best open general vision model in the table, Qwen3.8-Flash-Next at 125 billion parameters, reaches 55.05. A 4-billion-parameter model trained on 9,000 samples lands at 55.81 — above the 125B generalist, above Kimi-K3 at 2.8 trillion parameters, above GLM-5.3-Flash at 320 billion.

If that number survived contact with an independent harness it would be the story of the month. It does not, yet, for a reason the paper is upfront about. ZoomGainBench is constructed by the same pipeline that produced the training data, and the paper's own ablation table lists it under a column headed "Held-in." The questions are not the training questions, but they are drawn from the same distribution by the same machinery. A model fine-tuned on that distribution has an advantage that a model evaluated zero-shot does not, and the size of the advantage is exactly what nobody outside inclusionAI can measure today.

The held-out evidence is less spectacular and more convincing. The paper reports an average over ten held-out benchmarks: 54.20 for the untrained Qwen3.5-4B, 60.96 for the ZoomGain-trained 4B. A 6.76-point average gain on benchmarks the training never saw is a real result. It is just not the same claim as beating a 2.8-trillion-parameter model.

Worth noting too that the win is not uniform. Comparing the 4B against its own base across the paper's main table:

• Large gains. Visual Probe 43.30 to 57.28. O3-Bench 54.20 to 63.19. Zoom Bench 48.88 to 59.88. HR-Bench 4K 81.25 to 87.25. BabyVision 14.69 to 21.13.

• Moderate gains. MME-RealWorld-Lite 57.89 to 61.59. MMStar 66.80 to 72.53. HR-Bench 8K 77.63 to 83.88. MMVP 57.33 to 61.33.

• One regression. TreeBench-style "Tree" reads 49.38 before training and 49.14 after. A 0.24-point dip, well inside noise, but it is in the table rather than out of it, which is a good sign about the table.

Single-column scoreboard card titled ZoomGain-4B: what the repository says, with six rows: Parameters 4.54B, Parent model Qwen3.5-4B, ZoomGainBench 55.81 held-in, ten held-out benchmarks 60.96 versus 54.20, Hosted API none, Announcement none found, plus a footer noting the figures are vendor-reported and unaudited.

What is genuinely not confirmed

Four things, and they matter more than the usual disclaimer.

No independent evaluation exists. Not a single third-party number. The 55.81, the 60.96 average, the 90.25 that the 9B reaches on HR-Bench 4K — all of it is the authors measuring their own model on their own harness. That is normal for a paper and it is not evidence of anything being wrong. It does mean the numbers should be read as claims.

No hosted endpoint. Neither checkpoint lists an inference provider. There is no API, no playground, no rate card. If you want to use this model in the next month you are downloading 8.42 GB and running it yourself.

The name disagrees with the parameter count in three places. The model is called 4B, the GGUF card says 4B, the repository badge says 5B, and the actual safetensors total is 4.54B. The extra half-billion is the vision encoder and projector, which is the ordinary explanation — but a card that cannot keep one number consistent on day one has not been proofread, and that is a fair thing to weigh when you are deciding how much of the rest to trust.

No statement of intent. Will there be a 30B? A hosted version? A v2 with a held-out benchmark? Nothing in the repository says. The README's update log lists three lines, all dated 2026.11.11, all describing what was released rather than what comes next.

Screenshot of the inclusionAI/ZoomGain GitHub repository, captured 11 November 2026, showing the repository name, description, commit history and file listing at the top level.

Running it, today, without a vendor

The GGUF path is the practical one and it has a hard prerequisite: you need a llama.cpp build with Q​wen3.5 multimodal support. The card is explicit that the language model and the mmproj projector are separate files and that image input requires both. The CLI example passes -m for the model, --mmproj for the projector, --image for the picture and runs at temperature 0 with thinking disabled, which is also what the evaluation harness defaults to. llama-server works the same way on port 8080. At Q4_K_M you are holding 2.71 GB of language model plus a projector, which fits comfortably on a laptop; at F16 you are at 8.42 GB plus projector, which is a workstation or a single mid-range GPU.

For anything at volume, vLLM 0.19.1 with transformers 5.7.0 is the supported stack, and the repository's own evaluation script takes a MODEL_PATH, a GPUS list and a LANGUAGE switch that runs English only, Chinese only, or both for 2,000 tasks. That is enough scaffolding to reproduce the reported numbers on your own hardware, which is more than most quiet releases offer.

If you would rather not stand up a GPU to find out whether the approach is worth anything, the honest interim move is to prototype the crop-and-ask loop against a hosted vision model you can call today. Our catalogue carries Qwen3-VL-8B at $0.18 in and $0.70 out per million tokens, and the Q​wen3.5 line from qwen3.5-35b-a3b at $0.057 and $0.459 upward — one key, no per-vendor contract, and a routing rule that fails over to a second model if the first stalls. ZoomGain-4B itself is not on our catalogue and I would not pretend otherwise; the point of the hosted detour is to establish whether fine-detail questions are the thing your product actually gets wrong, before you spend a week on a self-hosted 4B that only answers them better than its own parent.

What would change this from a curiosity into a decision

Three signals, in order of how much they would tell you. First, an independent run of ZoomGainBench by someone with no stake in the pipeline — the benchmark is public, the harness is in the repository, and the barrier to entry is a GPU and an afternoon, so this is likely to happen before any of the others. Second, a held-out extension: if the authors answer the obvious objection by evaluating on a benchmark they had no hand in building, the 55.81 stops being a number about their own test set. Third, a hosted deployment, which would tell you the weights are actually servable at acceptable latency rather than merely downloadable.

Until one of those lands, the correct description of ZoomGain-4B is not "small vision model beats frontier". It is: a research group published a training recipe that lifts a 4B vision model by 6.76 points on average across ten benchmarks it did not train on, and simultaneously published a benchmark that shows the same trick producing a much larger number on questions built the same way. The first half is the finding. The second half is the marketing, and it arrived without a press release, which is at least a refreshing order for those two things to come in.