ZoomGain-4B: 4GB+ VRAM, GGUF Builds

ZoomGain-4B: 4GB+ VRAM, GGUF Builds

At a Glance

Item Value
Repository inclusionAI/ZoomGain-4B
Family guide ZoomGain guide (3 articles)
Publisher guide inclusionAI (Ant Group): models and licenses
Published 2026-11-10
Formats safetensors
Paper arXiv:2611.14076
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

inclusionAI has released “ZoomGain-4B", a vision-language model specialized in fine-grained visual recognition. The base architecture adopts the Qwen3.5 series, and it is distributed along with evaluation benchmarks and reinforcement learning datasets released at the same time.

According to the paper referenced by this model, “ZoomGain: Mining Global-to-Zoom Perception Gaps for Fine-Grained Visual Supervision" (arXiv:2611.14076), vision-language models have the characteristic that even fine details, which are difficult to recognize in the entire image, can sometimes be recognized by zooming in on the relevant area. ZoomGain was developed focusing on this “global-to-zoom perception gap," aiming to efficiently generate high-quality visual training data by automatically constructing questions about the entire image from detailed visual information obtained only through zooming, and to reinforcement-learn the model.

Specifications

  • Architecture: Qwen3_5ForConditionalGeneration (qwen3_5)

Performance

Evaluation results have been published in the model’s repository card and the abstract of the referenced paper. Although structured score tables are not embedded directly in the body of the model card and instead refer to tables in the paper (such as Table 3), the following quantitative and qualitative evaluation results are known from the published materials.

On the evaluation benchmark “ZoomGainBench" independently constructed by the publishers, it is reported that among the 28 evaluated vision-language models, the highest score remained at 61.08. This benchmark consists of 1,000 questions (totaling 2,000 tasks in English and Chinese) for 941 images, and evaluates tasks that strongly depend on local image details, such as attribute recognition, counting, and OCR (character recognition).

Regarding the learning results of the model, it is stated that training on just 9,000 ZoomGain reinforcement learning data significantly improved the visual recognition capabilities of the 4B and 9B models. Through this training, it is claimed to have shown results exceeding the performance of large models with hundred-billion-parameter scales on various recognition benchmarks. In addition, for its sibling model of the same lineage, ZoomGain-9B, it is reported to have recorded a high score of 90.25 on the HR-Bench 4K benchmark, achieving State-of-the-Art (SoTA) performance.

These results indicate that the approach of reinforcing and learning using information from zoomed regions for subtle clues that are hard to catch from the entire image has the effect of significantly boosting recognition accuracy even in relatively small 4B-class models. On the other hand, it is also suggested that challenging tasks in perfectly identifying fine visual details without misrecognition remain, as the highest score by existing 28 models on ZoomGainBench remained in the 61-point range and multiple-choice questions tend to easily lead to specific incorrect answer options. Note that these figures and evaluations are based on measurements by the publishers themselves.

Strengths and Use Cases

ZoomGain-4B specializes in tasks that capture fine-grained visual perception, which is easily overlooked from the entire image, and perform reasoning based on detailed clues.

The research underlying this model focuses on the phenomenon where details that ordinary vision-language models cannot distinguish in the entire image become recognizable by zooming in on the relevant location. According to the model card and paper abstract, ZoomGain is enhanced by data constructed through the following steps:

  1. Ground the target: Multiple models locate the target object within the entire image, and the region where consensus is obtained is cropped.
  2. Mine evidence: The cropped region is compared with the caption of the entire image, converting detailed information revealed for the first time by zooming into questions about the entire image.
  3. Verify questions and answers: Answers are cross-checked on the cropped image, eliminating ambiguous questions or those answerable without images based on language priors, and constructing distractors.

Because it is trained through such a process, this model shines in tasks that require actual fine visual evidence rather than relying on linguistic guesswork. As shown in the composition of the evaluation benchmark “ZoomGainBench", it is specifically suited for the following uses:

  • Attribute Recognition: Identifying detailed features such as fine colors, patterns, shapes, and states of the subject.
  • Counting: Accurately identifying and counting small objects or elements densely packed on the screen.
  • Local OCR: Accurately reading small text, signs, and document descriptions in images that tend to blur in reduced displays.

Furthermore, since the evaluation dataset is constructed to support both English and Chinese, application to multilingual fine-grained visual question answering is also anticipated.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 4.5B parameters

Your VRAM Quantization File size Est. memory needed
4GB (laptop iGPU / phone class) Q4_K_M 2.5GB 3.0GB
8GB (RTX 4060 / 3060 Ti, etc.) Q8_0 4.2GB 5.0GB
12GB (RTX 4070 / 3060 12GB, etc.) F16 7.8GB 9.4GB

Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: registered, vLLM: registered, MLX (mlx-lm): registered.

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. File sizes are measured from the converted build inclusionAI/ZoomGain-4B-GGUF. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

The publisher distributes this model as safetensors.

Compression: the Q4_K_M build measures 4.73 bits per weight — about 30% the size of the original 16-bit weights, calculated by this site from the actual file sizes.

Compiled by this site’s code from the published formats and the license field.

Our Own Measurements

Measurements We Did Not Take

This model does not state a license, so we cannot confirm permission to use it. Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).

How to Get It

Model weights are released in safetensors format and can be downloaded using the Hugging Face CLI. Since this model is not gated, it can be acquired directly without being prompted for individual agreement to terms of use.

The command to download the model weights is as follows:

hf download inclusionAI/ZoomGain-4B --local-dir "./ZoomGain-4B"

Building the inference and evaluation environment requires cloning the official repository and installing dependent packages. The repository specifies the introduction of vllm==0.19.1 and transformers==5.7.0 as inference engines and libraries.

git clone https://github.com/inclusionAI/ZoomGain.git
cd ZoomGain

python -m pip install -r eval_ZoomGainBench/requirements.txt \
  'vllm==0.19.1' 'transformers==5.7.0'

When running evaluations using ZoomGainBench in a local environment, instructions guide users to download the benchmark dataset (inclusionAI/ZoomGainBench) and run the evaluation script (eval_ZoomGainBench/eval.sh) within the repository.

Quantized and Converted Variants

→ Scroll horizontally to see all columns

Added Publisher Format Repository Smallest VRAM tier (build, est. memory)
2026-11-17 bartowski GGUF (imatrix) bartowski/ZoomGain-4B-GGUF Q6_K_S 4.0GB (fits in 4GB VRAM)
2026-11-17 inclusionAI GGUF inclusionAI/ZoomGain-4B-GGUF Q4_K_M 3.0GB (fits in 4GB VRAM)

File sizes of each build:

  • Available builds in bartowski/ZoomGain-4B-GGUF: IQ2_M 1.5GB / Q2_K 1.6GB / IQ3_XXS 1.8GB / Q3_K_S 1.8GB / IQ3_XS 1.8GB / Q3_K_M 1.9GB / Q3_K_L 2.0GB / IQ3_M 2.1GB / IQ4_XS 2.3GB / Q4_K_S 2.4GB / Q4_0 2.5GB / IQ4_NL 2.5GB / Q4_K_M 2.5GB / Q4_1 2.7GB / Q4_K_L 2.7GB / Q5_K_S 2.9GB / Q5_K_M 3.1GB / Q6_K_S 3.3GB / Q6_K 3.4GB / Q6_K_L 3.6GB / Q8_0 4.2GB / BF16 7.8GB
  • Available builds in inclusionAI/ZoomGain-4B-GGUF: Q4_K_M 2.5GB / Q8_0 4.2GB / F16 7.8GB

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.

Related Articles

What to Read Next

Sources

Update History

  • 2026-11-17: Added converted builds to “Quantized and Converted Variants”: bartowski/ZoomGain-4B-GGUF, inclusionAI/ZoomGain-4B-GGUF
  • 2026-11-17: Updated the hardware requirements table with the actual file sizes of inclusionAI/ZoomGain-4B-GGUF.
  • 2026-11-17: Changed the title to show what the article covers (VRAM requirements, file list, etc.).