# ZwZ2-4B

> ZwZ2-4B is the Qwen3.5-4B-based model in Zooming-without-Zooming-v2 (ZwZ2). It learns to connect visual evidence across regions through comparison, aggregation, matching, and multi-step reasoning. Training uses full images; inference requires no zooming tools.

- **Organisation:** inclusionAI
- **Category:** Vision-Language Models
- **Parameters:** 4.5B
- **Context length:** 262,144 tokens
- **Licence:** Unknown licence (Check licence)
- **Released:** 2026-10-15
- **Hugging Face:** 1.1K downloads (30 days), 12 likes

## Estimated memory

| Quant | ≈ GB |
|---|---|
| Q2_K | 2.96 |
| Q3_K_M | 3.67 |
| Q4_K_M | 4.19 |
| Q5_K_M | 4.66 |
| Q6_K | 5.16 |
| Q8_0 | 6.22 |
| F16 | 10.38 |

## Runs with

- [llama.cpp](https://selfhost.directory/project/llamacpp)
- [SGLang](https://selfhost.directory/project/sglang)
- [Transformers](https://github.com/huggingface/transformers)
- [vLLM](https://selfhost.directory/project/vllm)
- [Jan](https://jan.ai) (community)
- [KoboldCpp](https://selfhost.directory/project/koboldcpp) (community)
- [LM Studio](https://lmstudio.ai) (community)
- [LocalAI](https://selfhost.directory/project/localai) (community)
- [Ollama](https://selfhost.directory/project/ollama) (community)

## About

ZwZ2-4B is the Qwen3.5-4B-based model in Zooming-without-Zooming-v2 (ZwZ2). It learns to connect visual evidence across regions through comparison, aggregation, matching, and multi-step reasoning. Training uses full images; inference requires no zooming tools.

On 25 benchmarks across six domains, it improves over its base model on every benchmark, increasing overall accuracy from 49.41 to 54.79 (+5.38 points). Gains extend from fine-grained perception to spatial reasoning, long-context understanding, multimodal agents, and general visual reasoning.

EvidenceConnect builds verified multi-region questions from SA-1B images:

• Group 3–7 related regions and generate questions using privileged zoomed views.
• Check answers through cross-model agreement and drop-one region tests.
• Convert verified questions to four-option tasks and train with GRPO, without an additional supervised fine-tuning stage.

All three model sizes use approximately 20K training examples. The policy sees full images with colored region outlines; privileged crop layouts are used only for data construction. See ZwZ2-RL-VQA for the data schema and synthesis process.

Editable vector PDF. The figure shows eight representative benchmarks; the complete results follow.

## Alternatives

- [Qwen/Qwen3.5-4B](https://selfhost.directory/models/qwen/qwen3.5-4b) (4.7B, Apache 2.0, ≈4.26 GB at Q4): Over recent months, we have intensified our focus on developing foundation models that deliver exceptional utility and performance.
- [Qwen/Qwen3-VL-4B-Instruct](https://selfhost.directory/models/qwen/qwen3-vl-4b-instruct) (4.4B, Apache 2.0, ≈4.26 GB at Q4): Meet Qwen3-VL — the most powerful vision-language model in the Qwen series to date.
- [Qwen/Qwen3.5-4B-Base](https://selfhost.directory/models/qwen/qwen3.5-4b-base) (4.7B, Apache 2.0, ≈4.26 GB at Q4): Over recent months, we have intensified our focus on developing foundation models that deliver exceptional utility and performance.
- [h2oai/h2o-lightning-4b](https://selfhost.directory/models/h2oai/h2o-lightning-4b) (4.5B, Apache 2.0, ≈4.19 GB at Q4): H2O-Lightning-4B is a 4B-parameter decision model from H2O.ai, built on Qwen/Qwen3.5-4B. It answers typed decision questions about a record (a document, a ticket, a policy, a conversation) and about images that come with the record (photos, screenshots, scanned documents, charts).
- [tencent/WeVisOne-VL-4B](https://selfhost.directory/models/tencent/wevisone-vl-4b) (4.8B, Apache 2.0, ≈4.49 GB at Q4): WeVisOne-VL (WeChat Vision All-in-One) is a unified vision-language model developed by WeChat Vision, Tencent Inc. Built upon Qwen3-VL, it unifies object-level perception (open-vocabulary object detection, referring localization, object counting, object pointing, and OCR) and pixel-level perception…
- [google/medgemma-1.5-4b-it](https://selfhost.directory/models/google/medgemma-1.5-4b-it) (4.3B, health-ai-developer-foundations, ≈4.11 GB at Q4): * License: The use of MedGemma is governed by the Health AI Developer Foundations terms of use. MedGemma has not been evaluated or optimized for multi-turn applications.

More: https://selfhost.directory/models/inclusionai/zwz2-4b/alternatives

## Runs well on

- [GeForce RTX 4060](https://selfhost.directory/models/inclusionai/zwz2-4b/on/rtx-4060): Q8_0, est. 31 tok/s
- [Apple M4 (16 GB)](https://selfhost.directory/models/inclusionai/zwz2-4b/on/apple-m4-16gb): F16, est. 7.3 tok/s
- [Radeon RX 6700 XT](https://selfhost.directory/models/inclusionai/zwz2-4b/on/rx-6700-xt): F16, est. 23 tok/s
- [GeForce RTX 3060 12GB](https://selfhost.directory/models/inclusionai/zwz2-4b/on/rtx-3060): F16, est. 22 tok/s
- [Arc B580](https://selfhost.directory/models/inclusionai/zwz2-4b/on/arc-b580): F16, est. 28 tok/s
- [Tesla T4](https://selfhost.directory/models/inclusionai/zwz2-4b/on/tesla-t4): F16, est. 19 tok/s

## Similar size

- [Qwen/Qwen3.5-4B](https://selfhost.directory/models/qwen/qwen3.5-4b) (4.7B)
- [google/gemma-3-4b-it](https://selfhost.directory/models/google/gemma-3-4b-it) (4.3B)
- [Qwen/Qwen3-VL-4B-Instruct](https://selfhost.directory/models/qwen/qwen3-vl-4b-instruct) (4.4B)
- [Qwen/Qwen2.5-VL-3B-Instruct](https://selfhost.directory/models/qwen/qwen2.5-vl-3b-instruct) (3.8B)
- [google/translategemma-4b-it](https://selfhost.directory/models/google/translategemma-4b-it) (5B)
- [google/medgemma-1.5-4b-it](https://selfhost.directory/models/google/medgemma-1.5-4b-it) (4.3B)

## FAQ

**How much VRAM does ZwZ2-4B need?**
About 4.2 GB at Q4_K_M with an 8K context. Q8_0 needs about 6.2 GB and full precision about 10 GB. Longer contexts need more memory for the KV cache.

**Can ZwZ2-4B run on the RTX 4060?**
Yes. It fits at Q8_0 (about 5.9 GB), at an estimated 31 tokens/s. It fits comfortably on the Apple M4 (16 GB), Radeon RX 6700 XT, RTX 3060 12GB.

**Can I use ZwZ2-4B commercially?**
Unclear. It is released under Unknown licence. No licence stated: assume all rights reserved. Check the licence on the model card before using it in a product.

**What is the context length of ZwZ2-4B?**
262,144 tokens (256K). The full window needs far more memory than an 8K context.

**How do I run ZwZ2-4B with Ollama or llama.cpp?**
With Ollama: ollama run hf.co/inclusionAI/ZwZ2-4B-GGUF:Q4_K_M. With llama.cpp: llama-server -hf inclusionAI/ZwZ2-4B-GGUF:Q4_K_M, which starts an OpenAI-compatible server.

## Links

- huggingface: https://huggingface.co/inclusionAI/ZwZ2-4B
- github: https://github.com/inclusionAI/Zooming-without-Zooming
- paper: https://arxiv.org/abs/2602.11858

---
Source: https://selfhost.directory/models/inclusionai/zwz2-4b