# ZwZ2-9B

> ZwZ2-9B is the Qwen3.5-9B-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:** 9.4B
- **Context length:** 262,144 tokens
- **Licence:** Unknown licence (Check licence)
- **Released:** 2026-10-16
- **Hugging Face:** 1.5K downloads (30 days), 17 likes

## Estimated memory

| Quant | ≈ GB |
|---|---|
| Q2_K | 4.53 |
| Q3_K_M | 6 |
| Q4_K_M | 7.08 |
| Q5_K_M | 8.04 |
| Q6_K | 9.08 |
| Q8_0 | 11.28 |
| F16 | 19.9 |

## 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)

## About

ZwZ2-9B is the Qwen3.5-9B-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 51.63 to 56.59 (+4.96 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/Qwen3.5-9B](https://selfhost.directory/models/qwen/qwen3.5-9b) (9.7B, Apache 2.0, ≈7.22 GB at Q4): Over recent months, we have intensified our focus on developing foundation models that deliver exceptional utility and performance.
- [Qwen/Qwen3-VL-8B-Instruct](https://selfhost.directory/models/qwen/qwen3-vl-8b-instruct) (8.8B, Apache 2.0, ≈6.82 GB at Q4): Meet Qwen3-VL — the most powerful vision-language model in the Qwen series to date.
- [Zyphra/ZAYA1-VL-8B](https://selfhost.directory/models/zyphra/zaya1-vl-8b) (9.7B, Apache 2.0, ≈6.58 GB at Q4): ZAYA1-VL-8B is a vision-language model (VLM) built upon Zyphra's ZAYA1-8B LLM. It has state-of-the-art performance among VLMs for its size and inference efficiency.
- [Qwen/Qwen3.5-9B-Base](https://selfhost.directory/models/qwen/qwen3.5-9b-base) (9.7B, Apache 2.0, ≈7.22 GB at Q4): Over recent months, we have intensified our focus on developing foundation models that deliver exceptional utility and performance.
- [XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B](https://selfhost.directory/models/xiaomimimo/mimo-v2.6-distill-qwen-9b) (9.4B, MIT, ≈7.08 GB at Q4): MiMo-V2.6-Distill-Qwen-9B is a 9B agentic model developed by Xiaomi MiMo through supervised fine-tuning of Qwen3.5-9B on MiMo-generated data. It covers coding, general-purpose agent tasks, visual coding, and cybersecurity.
- [black-forest-labs/flux-3-image-klein-pup](https://selfhost.directory/models/black-forest-labs/flux-3-image-klein-pup) (9.4B, Apache 2.0, ≈7.08 GB at Q4): FLUX 3 Image Klein Prompt Upsampler is a 9B vision-language model that expands a short image request into the detailed prompt FLUX 3 Image Klein works best with. It takes the request, up to 10 reference images and an optional list of bounding boxes, and returns one prompt.

More: https://selfhost.directory/models/inclusionai/zwz2-9b/alternatives

## Runs well on

- [GeForce RTX 4060](https://selfhost.directory/models/inclusionai/zwz2-9b/on/rtx-4060): Q4_K_M, est. 26 tok/s
- [Apple M4 (16 GB)](https://selfhost.directory/models/inclusionai/zwz2-9b/on/apple-m4-16gb): Q6_K, est. 8.5 tok/s
- [Radeon RX 6700 XT](https://selfhost.directory/models/inclusionai/zwz2-9b/on/rx-6700-xt): Q8_0, est. 21 tok/s
- [GeForce RTX 3060 12GB](https://selfhost.directory/models/inclusionai/zwz2-9b/on/rtx-3060): Q8_0, est. 20 tok/s
- [Arc B580](https://selfhost.directory/models/inclusionai/zwz2-9b/on/arc-b580): Q8_0, est. 25 tok/s
- [Tesla T4](https://selfhost.directory/models/inclusionai/zwz2-9b/on/tesla-t4): Q8_0, est. 18 tok/s

## Similar size

- [Qwen/Qwen3.5-9B](https://selfhost.directory/models/qwen/qwen3.5-9b) (9.7B)
- [Qwen/Qwen3-VL-8B-Instruct](https://selfhost.directory/models/qwen/qwen3-vl-8b-instruct) (8.8B)
- [Qwen/Qwen2.5-VL-7B-Instruct](https://selfhost.directory/models/qwen/qwen2.5-vl-7b-instruct) (8.3B)
- [google/gemma-3-12b-it](https://selfhost.directory/models/google/gemma-3-12b-it) (12.2B)
- [XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B](https://selfhost.directory/models/xiaomimimo/mimo-v2.6-distill-qwen-9b) (9.4B)
- [meta-llama/Llama-3.2-11B-Vision-Instruct](https://selfhost.directory/models/meta-llama/llama-3.2-11b-vision-instruct) (10.7B)

## FAQ

**How much VRAM does ZwZ2-9B need?**
About 7.1 GB at Q4_K_M with an 8K context. Q8_0 needs about 11 GB and full precision about 20 GB. Longer contexts need more memory for the KV cache.

**Can ZwZ2-9B run on the RTX 4060?**
Yes. It fits at Q4_K_M (about 7.0 GB), at an estimated 26 tokens/s. It fits comfortably on the Apple M4 (16 GB), Radeon RX 6700 XT, RTX 3060 12GB.

**Can I use ZwZ2-9B 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 ZwZ2-9B?**
262,144 tokens (256K). The full window needs far more memory than an 8K context.

**How do I run ZwZ2-9B with llama.cpp?**
Run llama-server -hf inclusionAI/ZwZ2-9B-GGUF:Q4_K_M. It downloads the GGUF and starts an OpenAI-compatible server.

## Links

- huggingface: https://huggingface.co/inclusionAI/ZwZ2-9B
- github: https://github.com/inclusionAI/Zooming-without-Zooming
- paper: https://arxiv.org/abs/2602.11858

---
Source: https://selfhost.directory/models/inclusionai/zwz2-9b