ZwZ2-4B-GGUF Vision-Language Model: 4GB+ VRAM

ZwZ2-4B-GGUF Vision-Language Model: 4GB+ VRAM

At a Glance

Item Value
Repository inclusionAI/ZwZ2-4B-GGUF
Family guide Qwen3.5 guide (8 articles)
Publisher guide inclusionAI (Ant Group): models and licenses
Published 2026-10-30
License apache-2.0
Formats GGUF
Paper arXiv:2610.37037
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 GGUF-format weights for the open-weight vision-language model “ZwZ2-4B". This model was published as part of the “Zooming without Zooming v2 (ZwZ2)" project, adopting “Qwen3.5-4B" as its base model.

According to the paper cited in the model card, “Zooming without Zooming v2: From Seeing Details to Connecting Evidence for Generalizable Multimodal Perception" (arXiv:2610.37037), the model was developed to improve the capability of connecting evidence—comparing, aggregating, and verifying visual evidence across multiple regions within an image through multi-step reasoning. It is designed to capture and reason about fine-grained visual information from the entire image without requiring external zooming tools during inference. The newly released repository is a converted version in GGUF format that can be run with llama.cpp and others, provided in two separate components: the language model portion and the vision encoder-projector (mmproj).

Specifications

The specifications listed in the original model card and public repository are as follows:

  • Base model: Qwen3.5-4B
  • Architecture: Qwen3_5ForConditionalGeneration
  • Model configuration: Split configuration consisting of a language model component and a vision encoder-projector component (mmproj)
  • Available precisions/quantization types: FP16, Q8_0, Q4_K_M

Performance

Since the newly released repository is a quantized GGUF version, the benchmark scores introduced below are based on the measurement results stated in the model card of the source “ZwZ2-4B" model. Please note that quantization processing may cause some degradation in accuracy compared to the original model.

The original model card shows the following correspondence relationship across the model series:

Model Base model
ZwZ2-4B Qwen3.5-4B
ZwZ2-9B Qwen3.5-9B
ZwZ2-27B Qwen3.6-27B

According to the original model card, it is reported that ZwZ2-4B outperformed the scores of the base Qwen3.5-4B across all 25 benchmarks spanning 6 domains. The overall average accuracy improved by 5.38 points, rising from 49.41% for the base model to 54.79%.

It is stated that this improvement extends across a wide range of areas, not only limited to fine-grained perception, but also covering spatial reasoning, long-context understanding, multimodal agents, and general visual reasoning. The training utilized approximately 20,000 multi-region question data generated from SA-1B images using “EvidenceConnect", and reinforcement learning via GRPO (Group Relative Policy Optimization) was applied without an intermediate supervised fine-tuning (SFT) stage.

Note that these evaluation figures are results measured and published by the model creators themselves, rather than verified by a third-party organization. Furthermore, the detailed numerical tables for individual benchmarks are presented in chart format within the main model card, and specific breakdown figures for each task are limited to the information described on the card.

Strengths and Use Cases

What “ZwZ2-4B" excels at are advanced visual reasoning tasks where answers are derived by cross-referencing multiple regions present within a single image while aggregating and verifying visual evidence.

While conventional vision-language models sometimes employ methods of cropping or zooming in on a specific part of an image to grasp details, this model does not require external zooming tools during inference. It is designed to achieve both fine-grained perception and broad spatial reasoning while maintaining the overall composition of the image.

Based on descriptions in the model card and the paper, it is suited for the following scenarios and use cases:

  • Comparison and verification of multiple locations within an image: Suitable for processing that grasps the placement, shape, color, and mutual relationships of multiple objects or regions located in different parts of the screen, connecting them logically.
  • Multimodal spatial reasoning: Effective in tasks that decode relative positional relationships between objects or complex geometric and physical layouts.
  • Understanding long-context and lengthy visual information: Handles long-context tasks that simultaneously track numerous elements within an image and arrive at answers through multi-step processes.
  • Multimodal agents and embodied AI: Anticipated for applications in agent processing where surrounding environment images are surveyed to accurately recognize multiple targets and determine subsequent actions.
  • Image-based dialogue: Supports general conversational tasks where questions about image content are answered with detailed explanations.

Additionally, the supported language tags include English (en), primarily anticipating instructions and conversations in English.

How It Differs from Similar Models

In past articles, this site has covered “Tev1-4B-experimental" 4B model specialized for structured data decision-making: Required VRAM 4GB~, which adopted the same “Qwen3.5-4B" as its base model.

While both models are built upon the 4B-scale Qwen3.5 family, their intended objectives and training approaches differ significantly. “Tev1-4B-experimental" is an experimental model specialized in decision-making tasks that return a single option label from structured text inputs (states, questions, and choices), pursuing efficiency as a classifier through supervised fine-tuning (SFT).

In contrast, the “ZwZ2-4B" covered here is a vision-language model (VLM) combining images and text, specialized in combining visual evidence spanning multiple regions within an image. It is also distinctive in that it was trained via reinforcement learning using GRPO (Group Relative Policy Optimization) without an SFT stage. While Tev1 aims for text-centric decision-making classification, ZwZ2-4B focuses on reading spatial and detailed information from the entire image to perform advanced visual reasoning.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 4.5B parameters (taken from the base model inclusionAI/ZwZ2-4B)

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

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. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

Runs in Ollama, LM Studio and llama.cpp as-is.

It is distributed in GGUF, so no conversion is needed.

License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.

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, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Recent Models in the Same Size Class

Models with 4–15B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.

→ Scroll horizontally to see all columns

Model Parameters Smallest VRAM tier License Our article
tencent/Youtu-Parsing-Omni 5.3B 12GB youtu-parsing Youtu-Parsing-Omni Omnimodal Analysis Model: 12GB+ VRAM (2026-11-17)
nvidia/Muse-Spark-Open-816B-A42B-DFlash2 4.4B 12GB nvidia-open-model-agreement-license Muse-Spark-Open-816B-A42B-DFlash2 Vision-Language Model: 12GB+ VRAM (2026-11-17)
tencent/WeVisOne-VL-4B 4.8B 4GB apache-2.0 Tencent Releases WeVisOne-VL Vision Language Models (2026-11-17)
inclusionAI/ZwZ2-9B-GGUF 9.4B 8GB apache-2.0 ZwZ2-9B-GGUF Vision-Language Model: 8GB+ VRAM (2026-11-17)
microsoft/Mage-VL-Cowork 4.7B 12GB apache-2.0 Mage-VL-Cowork Vision-Language Model: 12GB+ VRAM, GGUF Builds (2026-11-17)

How to Get It

“ZwZ2-4B-GGUF" can be obtained directly from its Hugging Face repository (inclusionAI/ZwZ2-4B-GGUF). It is released under the Apache-2.0 license as an open repository that does not require pre-screening or agreement to terms of use.

This model is distributed in GGUF format and is provided split into two parts: the language model body and the component responsible for the vision encoder and projector (mmproj). Precisions of “FP16", “Q8_0", and “Q4_K_M" are available for each, allowing users to combine them freely according to their use case.

As for the execution environment, recent builds of “llama.cpp" that support Qwen3.5 multimodal features are specified. When running inference with image inputs, you must specify the corresponding mmproj file alongside the main language model.

When running inference from the CLI, use llama-mtmd-cli as specified below (the following is an example using Q8_0 for both the language model and vision components).

llama-mtmd-cli \
  -m ZwZ2-4B-Q8_0.gguf \
  --mmproj mmproj-ZwZ2-4B-Q8_0.gguf \
  --image /path/to/image.jpg \
  -p "Describe this image." \
  --temp 0 -n 2048

Additionally, if you wish to run it as an OpenAI-compatible API server or Web UI, you can launch a local server using llama-server.

llama-server \
  -m ZwZ2-4B-Q8_0.gguf \
  --mmproj mmproj-ZwZ2-4B-Q8_0.gguf \
  --host 127.0.0.1 --port 8080

If you want to use other precisions (such as FP16 or Q4_K_M), modify the specified file names accordingly when executing.

Related Articles

What to Read Next

Sources