ZoomGain-9B Fine-Grained Visual Language Model: 8GB+ VRAM, GGUF Builds

ZoomGain-9B Fine-Grained Visual Language Model: 8GB+ VRAM, GGUF Builds

At a Glance

Item Value
Repository inclusionAI/ZoomGain-9B
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 the vision-language model (VLM) “ZoomGain-9B".

According to the abstract of the paper “ZoomGain: Mining Global-to-Zoom Perception Gaps for Fine-Grained Visual Supervision" (https://arxiv.org/abs/2611.14076), this model was developed using the “ZoomGain" method, which leverages perception gaps between global images and zoomed regions to automatically generate and verify questions about global images based on visual details that only become apparent when zoomed in. It was created with the aim of enhancing the model’s minute visual recognition capabilities through reinforcement learning using 9,000 (9K) ZoomGain data points without manual question-answering edits.

Specifications

  • Parameters: 9.4B
  • Architecture: Qwen3_5ForConditionalGeneration

Performance

According to the model card and published values in the paper, ZoomGain-9B demonstrates high performance across multiple fine-grained visual recognition benchmarks and general visual recognition tests.

Specifically, it is reported to have achieved a score of 90.25 on the HR-Bench 4K benchmark. The paper’s abstract explains that 4B and 9B models trained on just 9K ZoomGain samples outperformed tens-of-billions-parameter scale models across various recognition benchmarks.

On the other hand, measurement results on the evaluation benchmark “ZoomGainBench" created by the development team consisting of 1,000 questions showed that the highest score among all 28 evaluated VLMs remained at 61.08, and challenges were also analyzed, such as the tendency for visual models to easily select specific choices (distractors) when giving incorrect answers.

Note that the scores listed are measured values independently published by the paper’s authors and developers, and are not figures verified independently by third parties.

Strengths and Use Cases

ZoomGain-9B is well-suited for tasks that accurately read and interpret minute visual information that is difficult to grasp from just a global view of an image, or detailed evidence that can only be confirmed by zooming in.

Specifically, it demonstrates strengths in advanced fine-grained visual tasks such as attribute recognition (determining detailed features of objects and people in images), counting (accurately counting crowded scenes or small objects), and OCR (optical character recognition) for reading small text in complex images. These are areas that many conventional vision-language models (VLMs) have commonly struggled with, and the model proves effective in situations where accurate answers based on actual visual evidence are required without relying on overall context or linguistic preconceptions.

Furthermore, the question and answer texts support both English and Chinese (bilingual), making it applicable to fine-grained Visual Question Answering in multilingual environments. It is expected to be used in a wide range of applications requiring high-precision visual backing, such as the analysis of complex visual materials, detailed text decoding of documents and labels, inspection of minute parts, and various agent processes that extract specific small-area information from wide-range images.

In addition, this model is suitable not only for inference and use, but also as a basis for research and development of reinforcement learning (RL) and data pipelines that enhance fine-grained visual recognition capabilities. Along with the release of this model, the development team has also published “ZoomGain-RL-Data", consisting of 9,000 training records automatically generated and verified utilizing the perception gap between global and zoomed regions, and “ZoomGainBench", an evaluation benchmark composed of 1,000 questions (2,000 tasks). By utilizing these, it is possible to carry out additional reinforcement learning on existing models or independently verify fine-grained recognition functions.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 9.4B parameters

Your VRAM Quantization File size Est. memory needed
8GB (RTX 4060 / 3060 Ti, etc.) Q4_K_M 5.2GB 6.3GB
12GB (RTX 4070 / 3060 12GB, etc.) Q8_0 8.9GB 10.6GB
24GB (RTX 4090 / 3090, etc.) F16 16.7GB 20.0GB

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-9B-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 4.75 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

This model is distributed in safetensors format. It is publicly available on the Hugging Face Hub, and account license agreement procedures (gated models) are not required for use.

To download the model weights to a local environment using the Hugging Face CLI, run the following command:

hf download inclusionAI/ZoomGain-9B --local-dir "./ZoomGain-9B"

As for operation and inference environments, vllm (version 0.19.1) and transformers (version 5.7.0) are specified as supported packages.

If you want to run evaluation work or your own reinforcement learning pipeline, you can obtain the source code from the official GitHub repository (https://github.com/inclusionAI/ZoomGain.git). An example procedure for setting up the environment is as follows:

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

export ASSET_DIR="$PWD/zoomgain-assets"
mkdir -p "$ASSET_DIR"
python -m pip install -U huggingface_hub

The procedure for installing dependency libraries when introducing the inference and “ZoomGainBench" evaluation environments is presented as follows:

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

If you want to independently try reinforcement learning (such as GRPO training using EasyR1), additional libraries such as a CUDA environment, PyTorch 2.10.0, and flash-attn==2.8.3 are required.

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-9B-GGUF IQ2_M 4.0GB (fits in 4GB VRAM)
2026-11-17 inclusionAI GGUF inclusionAI/ZoomGain-9B-GGUF Q4_K_M 6.3GB (fits in 8GB VRAM)

File sizes of each build:

  • Available builds in bartowski/ZoomGain-9B-GGUF: IQ2_M 3.3GB / Q2_K 3.4GB / IQ3_XXS 3.9GB / Q3_K_S 4.0GB / IQ3_XS 4.0GB / Q3_K_M 4.2GB / Q3_K_L 4.3GB / IQ3_M 4.5GB / IQ4_XS 4.9GB / Q4_0 5.1GB / Q4_K_S 5.1GB / IQ4_NL 5.4GB / Q4_K_M 5.4GB / Q4_1 5.5GB / Q4_K_L 5.8GB / Q5_K_S 6.0GB / Q5_K_M 6.4GB / Q6_K_S 7.0GB / Q6_K 7.3GB / Q6_K_L 7.5GB / Q8_0 8.9GB / BF16 16.7GB
  • Available builds in inclusionAI/ZoomGain-9B-GGUF: Q4_K_M 5.2GB / Q8_0 8.9GB / F16 16.7GB

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-9B-GGUF, inclusionAI/ZoomGain-9B-GGUF
  • 2026-11-17: Updated the hardware requirements table with the actual file sizes of inclusionAI/ZoomGain-9B-GGUF.
  • 2026-11-17: Changed the title to show what the article covers (VRAM requirements, file list, etc.).